WEBVTT 00:00:00.000 --> 00:00:00.958 align:middle line:90% 00:00:00.958 --> 00:00:02.395 align:middle line:90% [SQUEAKING] 00:00:02.395 --> 00:00:04.311 align:middle line:90% [RUSTLING] 00:00:04.311 --> 00:00:06.706 align:middle line:90% [CLICKING] 00:00:06.706 --> 00:00:09.365 align:middle line:90% 00:00:09.365 --> 00:00:10.740 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: OK. 00:00:10.740 --> 00:00:14.640 align:middle line:84% Let's get started right on time, both to set a good precedent, 00:00:14.640 --> 00:00:18.270 align:middle line:84% and there's quite a bit to do today. 00:00:18.270 --> 00:00:20.150 align:middle line:84% Oh, I will put that back down for one 00:00:20.150 --> 00:00:21.800 align:middle line:84% second so that those who just got in 00:00:21.800 --> 00:00:24.520 align:middle line:90% can see the agenda for today. 00:00:24.520 --> 00:00:28.400 align:middle line:90% 00:00:28.400 --> 00:00:31.020 align:middle line:84% First two things will just be like a minute or two, 00:00:31.020 --> 00:00:35.360 align:middle line:84% and then we'll spend the rest of the time on what everybody 00:00:35.360 --> 00:00:38.390 align:middle line:84% thinks is the coolest thing in the world 00:00:38.390 --> 00:00:41.030 align:middle line:84% these days in computation, artificial intelligence 00:00:41.030 --> 00:00:42.530 align:middle line:90% and machine learning. 00:00:42.530 --> 00:00:44.720 align:middle line:90% But this is not a class in that. 00:00:44.720 --> 00:00:48.150 align:middle line:84% We'll spend most of our time on feature extraction, 00:00:48.150 --> 00:00:52.293 align:middle line:84% the necessary evil before doing great things 00:00:52.293 --> 00:00:53.960 align:middle line:84% with artificial intelligence and machine 00:00:53.960 --> 00:00:57.530 align:middle line:84% learning when you're dealing with data that's not already 00:00:57.530 --> 00:00:58.880 align:middle line:90% numbers. 00:00:58.880 --> 00:01:04.010 align:middle line:84% I want to do a really quick discussion on something 00:01:04.010 --> 00:01:06.180 align:middle line:90% on problem set 7/8. 00:01:06.180 --> 00:01:08.130 align:middle line:84% I'm still listening to the compositions. 00:01:08.130 --> 00:01:10.730 align:middle line:84% Some of them are really, really cool. 00:01:10.730 --> 00:01:14.280 align:middle line:84% Just, they're all great, that I've heard so far. 00:01:14.280 --> 00:01:17.060 align:middle line:90% But some really went into it. 00:01:17.060 --> 00:01:21.200 align:middle line:84% But there was one thing that, among people 00:01:21.200 --> 00:01:23.960 align:middle line:84% who attempted every problem, ended up 00:01:23.960 --> 00:01:26.690 align:middle line:90% being a consistent error. 00:01:26.690 --> 00:01:28.250 align:middle line:90% Actually, two things. 00:01:28.250 --> 00:01:34.673 align:middle line:84% One that a lot of people missed here was in-- 00:01:34.673 --> 00:01:36.590 align:middle line:84% let's say you were in C major, because that'll 00:01:36.590 --> 00:01:38.670 align:middle line:90% make life easier. 00:01:38.670 --> 00:01:46.010 align:middle line:84% I gave some chords that were like this. 00:01:46.010 --> 00:01:50.030 align:middle line:84% And what do we call this in C major? 00:01:50.030 --> 00:01:52.110 align:middle line:84% What's the sca-- in Roman numeral. 00:01:52.110 --> 00:01:55.222 align:middle line:84% We're doing the Roman numeral exercise for a second. 00:01:55.222 --> 00:01:56.953 align:middle line:90% What scale degree? 00:01:56.953 --> 00:01:57.660 align:middle line:90% STUDENT: A ii. 00:01:57.660 --> 00:01:58.480 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: ii? 00:01:58.480 --> 00:01:58.980 align:middle line:90% Good. 00:01:58.980 --> 00:02:01.417 align:middle line:84% And what is the quality of the chord? 00:02:01.417 --> 00:02:02.250 align:middle line:90% STUDENT: Diminished. 00:02:02.250 --> 00:02:04.250 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: Diminished, great. 00:02:04.250 --> 00:02:07.443 align:middle line:84% So you don't have your problem set back for two reasons. 00:02:07.443 --> 00:02:09.610 align:middle line:84% One, there are a couple more songs I'm listening to. 00:02:09.610 --> 00:02:13.103 align:middle line:84% And also-- so how would we write ii diminished? 00:02:13.103 --> 00:02:13.770 align:middle line:90% STUDENT: Circle. 00:02:13.770 --> 00:02:15.853 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: With a circle, yeah. 00:02:15.853 --> 00:02:25.320 align:middle line:84% And then I noticed that my thing said, or if it's vii diminished, 00:02:25.320 --> 00:02:26.980 align:middle line:90% write this. 00:02:26.980 --> 00:02:31.150 align:middle line:84% And I did not say, in general, if it's diminished, write this. 00:02:31.150 --> 00:02:33.930 align:middle line:84% So I think that some people seem to write 00:02:33.930 --> 00:02:39.070 align:middle line:84% specific code for making sure that it only happened on vii. 00:02:39.070 --> 00:02:42.870 align:middle line:84% But ii diminished happens quite often. 00:02:42.870 --> 00:02:45.190 align:middle line:90% How does it happen in C major? 00:02:45.190 --> 00:02:49.770 align:middle line:84% You're borrowing from C minor for a little bit. 00:02:49.770 --> 00:02:55.020 align:middle line:84% So these borrowed chords, either flat and natural. 00:02:55.020 --> 00:03:00.348 align:middle line:84% And then you end up with that D-F-A flat chord. 00:03:00.348 --> 00:03:02.640 align:middle line:84% So that's where that comes in, but I want to make sure. 00:03:02.640 --> 00:03:04.560 align:middle line:84% I'm going to go back because there 00:03:04.560 --> 00:03:06.170 align:middle line:84% was an ambiguity in the instructions, 00:03:06.170 --> 00:03:09.960 align:middle line:84% and I want to make sure that everybody gets the little point 00:03:09.960 --> 00:03:13.360 align:middle line:84% back who made that, who didn't put that in. 00:03:13.360 --> 00:03:16.810 align:middle line:84% So that was one of the two systematic issues, 00:03:16.810 --> 00:03:18.820 align:middle line:84% and that came from an ambiguity for me. 00:03:18.820 --> 00:03:26.780 align:middle line:84% Here is the other one that came up a lot that I want to address. 00:03:26.780 --> 00:03:27.280 align:middle line:90% Oh, yeah. 00:03:27.280 --> 00:03:29.610 align:middle line:90% So we just went from C major. 00:03:29.610 --> 00:03:31.650 align:middle line:90% [PLAYS PIANO] 00:03:31.650 --> 00:03:32.550 align:middle line:90% And we played-- 00:03:32.550 --> 00:03:33.630 align:middle line:90% [PLAYS PIANO] 00:03:33.630 --> 00:03:36.750 align:middle line:90% --that ii diminished sound. 00:03:36.750 --> 00:03:40.490 align:middle line:90% Then what about this sound? 00:03:40.490 --> 00:03:42.990 align:middle line:90% [PLAYS PIANO] 00:03:42.990 --> 00:03:45.990 align:middle line:90% 00:03:45.990 --> 00:03:50.430 align:middle line:84% Anyone identify that chord by ear? 00:03:50.430 --> 00:03:51.918 align:middle line:90% Or just the-- what's that? 00:03:51.918 --> 00:03:52.710 align:middle line:90% STUDENT: Augmented. 00:03:52.710 --> 00:03:54.668 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: Augmented, great. 00:03:54.668 --> 00:03:56.050 align:middle line:90% [PLAYS PIANO] 00:03:56.050 --> 00:03:56.550 align:middle line:90% Good. 00:03:56.550 --> 00:04:03.160 align:middle line:84% I was in, doing that borrowing from minor thing. 00:04:03.160 --> 00:04:07.590 align:middle line:84% And I left my blue pen, which probably shows up better. 00:04:07.590 --> 00:04:10.310 align:middle line:84% I was borrowing from minor and I started on the third scale 00:04:10.310 --> 00:04:10.810 align:middle line:90% degree. 00:04:10.810 --> 00:04:13.420 align:middle line:84% And so this was not part of your Roman numeral thing, 00:04:13.420 --> 00:04:16.270 align:middle line:84% but it came on the roots question. 00:04:16.270 --> 00:04:18.959 align:middle line:90% 00:04:18.959 --> 00:04:23.975 align:middle line:90% What is the root of that chord? 00:04:23.975 --> 00:04:25.350 align:middle line:84% STUDENT: It could be any of them. 00:04:25.350 --> 00:04:26.450 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: No, it cannot. 00:04:26.450 --> 00:04:28.060 align:middle line:90% Ah, we'll get to that. 00:04:28.060 --> 00:04:29.860 align:middle line:90% That's where something happens. 00:04:29.860 --> 00:04:30.888 align:middle line:90% Yep, exactly. 00:04:30.888 --> 00:04:31.930 align:middle line:90% It's a great, great idea. 00:04:31.930 --> 00:04:32.745 align:middle line:90% John? 00:04:32.745 --> 00:04:33.090 align:middle line:90% STUDENT: E flat. 00:04:33.090 --> 00:04:34.632 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: E flat. 00:04:34.632 --> 00:04:38.020 align:middle line:84% Now, we'll get to it in just a second. 00:04:38.020 --> 00:04:46.100 align:middle line:90% What is the root of that chord? 00:04:46.100 --> 00:04:49.200 align:middle line:90% 00:04:49.200 --> 00:04:50.404 align:middle line:90% STUDENT: G. 00:04:50.404 --> 00:04:56.020 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: G. And let's see if I can make one. 00:04:56.020 --> 00:04:57.555 align:middle line:90% We'll go back to our E flat. 00:04:57.555 --> 00:05:02.320 align:middle line:90% 00:05:02.320 --> 00:05:06.123 align:middle line:90% What is the root of that one? 00:05:06.123 --> 00:05:06.790 align:middle line:90% STUDENT: C flat. 00:05:06.790 --> 00:05:08.332 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: C flat. 00:05:08.332 --> 00:05:09.730 align:middle line:90% Now, why is it? 00:05:09.730 --> 00:05:13.010 align:middle line:84% And so, just so we all have it in our ears again-- 00:05:13.010 --> 00:05:15.485 align:middle line:90% [PLAYS PIANO] 00:05:15.485 --> 00:05:17.470 align:middle line:90% 00:05:17.470 --> 00:05:20.750 align:middle line:84% Three chords, three exact same sound. 00:05:20.750 --> 00:05:23.410 align:middle line:84% But roots, the concept of root is 00:05:23.410 --> 00:05:27.640 align:middle line:84% something that comes from a triadic, sort 00:05:27.640 --> 00:05:30.170 align:middle line:90% of a spelling-enabled thing. 00:05:30.170 --> 00:05:32.900 align:middle line:90% So where do we find the root? 00:05:32.900 --> 00:05:35.780 align:middle line:84% We talked about the algorithm last time. 00:05:35.780 --> 00:05:37.100 align:middle line:90% What do we want to find? 00:05:37.100 --> 00:05:38.583 align:middle line:90% A bunch of stacked-- 00:05:38.583 --> 00:05:39.250 align:middle line:90% STUDENT: Thirds. 00:05:39.250 --> 00:05:40.100 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: Thirds. 00:05:40.100 --> 00:05:40.820 align:middle line:90% Good, I heard it. 00:05:40.820 --> 00:05:41.690 align:middle line:90% So stacked thirds. 00:05:41.690 --> 00:05:44.140 align:middle line:84% So here, we have a third, followed by a third, 00:05:44.140 --> 00:05:45.550 align:middle line:84% or a third and a fifth, depending 00:05:45.550 --> 00:05:46.760 align:middle line:90% on how you're counting. 00:05:46.760 --> 00:05:48.517 align:middle line:90% But here, what's this interval? 00:05:48.517 --> 00:05:49.350 align:middle line:90% STUDENT: Diminished. 00:05:49.350 --> 00:05:49.900 align:middle line:90% STUDENT: Diminished iv. 00:05:49.900 --> 00:05:50.410 align:middle line:90% STUDENT: Diminished iv. 00:05:50.410 --> 00:05:51.290 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: Diminished iv. 00:05:51.290 --> 00:05:52.735 align:middle line:90% Yeah, a iv of some sort. 00:05:52.735 --> 00:05:58.750 align:middle line:84% So we have to move it up the octave in order to get 00:05:58.750 --> 00:06:01.990 align:middle line:90% our stacked iii's. 00:06:01.990 --> 00:06:08.470 align:middle line:84% And here, we can move them both up an octave in order to get 00:06:08.470 --> 00:06:09.380 align:middle line:90% our stacked iii's. 00:06:09.380 --> 00:06:11.050 align:middle line:90% So the diminished-- 00:06:11.050 --> 00:06:14.230 align:middle line:84% The augmented triad is one of the places 00:06:14.230 --> 00:06:20.150 align:middle line:84% where the concept of root has everything to do with spelling. 00:06:20.150 --> 00:06:24.340 align:middle line:84% There's one other very common chord where the concept of root 00:06:24.340 --> 00:06:26.410 align:middle line:84% has everything to do with spelling. 00:06:26.410 --> 00:06:31.810 align:middle line:84% You saw a big smile, and it's not a triad. 00:06:31.810 --> 00:06:32.430 align:middle line:90% Go ahead. 00:06:32.430 --> 00:06:33.430 align:middle line:90% STUDENT: Diminished vii? 00:06:33.430 --> 00:06:35.470 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: Diminished vii chord, yeah. 00:06:35.470 --> 00:06:37.245 align:middle line:84% Here, I'll just keep it in the light. 00:06:37.245 --> 00:06:38.120 align:middle line:90% Assuming that's good. 00:06:38.120 --> 00:06:40.070 align:middle line:90% So yeah, diminished vii chord. 00:06:40.070 --> 00:06:42.370 align:middle line:90% Let's say we're in C minor. 00:06:42.370 --> 00:06:48.670 align:middle line:84% We're going to start on B. Would somebody shout 00:06:48.670 --> 00:06:50.750 align:middle line:90% out the next note? 00:06:50.750 --> 00:06:51.250 align:middle line:90% STUDENT: D. 00:06:51.250 --> 00:06:53.240 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: D. Keep going. 00:06:53.240 --> 00:06:53.740 align:middle line:90% STUDENT: F. 00:06:53.740 --> 00:06:55.630 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: F. And then? 00:06:55.630 --> 00:06:56.320 align:middle line:90% STUDENT: A flat. 00:06:56.320 --> 00:06:58.112 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: A flat, yeah. 00:06:58.112 --> 00:07:02.140 align:middle line:84% So vii, then we have vii with a root on B. 00:07:02.140 --> 00:07:07.000 align:middle line:84% But we can change this to move it up an octave 00:07:07.000 --> 00:07:08.796 align:middle line:90% and flip the enharmonic. 00:07:08.796 --> 00:07:12.610 align:middle line:90% 00:07:12.610 --> 00:07:21.100 align:middle line:84% And suddenly, we have iii, iii, iii, and our root becomes D. 00:07:21.100 --> 00:07:22.940 align:middle line:84% We've got a lot more to get through today, 00:07:22.940 --> 00:07:26.440 align:middle line:84% so I will leave it to a test to prove to yourself that we 00:07:26.440 --> 00:07:30.940 align:middle line:84% can keep doing this somehow, and get all, every one of the four 00:07:30.940 --> 00:07:37.420 align:middle line:84% notes, as a diminished-- as having its own, 00:07:37.420 --> 00:07:38.810 align:middle line:90% that note as a root. 00:07:38.810 --> 00:07:41.440 align:middle line:84% So each of the chord factors, depending on how you spell. 00:07:41.440 --> 00:07:43.190 align:middle line:84% And that's one of the reasons, by the way, 00:07:43.190 --> 00:07:45.490 align:middle line:84% this gets away from the computational side, but maybe 00:07:45.490 --> 00:07:48.080 align:middle line:84% a cool computational compositional tool. 00:07:48.080 --> 00:07:51.260 align:middle line:84% One of the reasons why a diminished vii chord is often 00:07:51.260 --> 00:07:55.290 align:middle line:84% a very useful way to modulate to remote keys, 00:07:55.290 --> 00:07:57.980 align:middle line:84% because you just sort of start emphasizing a different note 00:07:57.980 --> 00:07:59.005 align:middle line:90% as your root. 00:07:59.005 --> 00:08:02.021 align:middle line:90% 00:08:02.021 --> 00:08:05.545 align:middle line:84% Any questions on this, or on other things? 00:08:05.545 --> 00:08:08.390 align:middle line:84% We're all eager to get to feature extraction? 00:08:08.390 --> 00:08:16.490 align:middle line:84% On this, there should be a notebook that says, under today, 00:08:16.490 --> 00:08:21.350 align:middle line:90% 2023-ai-template-blank. 00:08:21.350 --> 00:08:22.250 align:middle line:90% That one. 00:08:22.250 --> 00:08:23.130 align:middle line:90% Is it? 00:08:23.130 --> 00:08:23.630 align:middle line:90% STUDENT: Yep, it's there. 00:08:23.630 --> 00:08:24.210 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: It's there. 00:08:24.210 --> 00:08:24.950 align:middle line:90% Good, good, good. 00:08:24.950 --> 00:08:28.580 align:middle line:84% So feel free to download that and put that where you are, 00:08:28.580 --> 00:08:30.990 align:middle line:90% and run it. 00:08:30.990 --> 00:08:32.549 align:middle line:90% It will install. 00:08:32.549 --> 00:08:35.120 align:middle line:84% If your hard drive is almost completely full, 00:08:35.120 --> 00:08:40.610 align:middle line:84% it will install a learning toolkit called Orange, which 00:08:40.610 --> 00:08:43.039 align:middle line:84% also installs another machine learning toolkit 00:08:43.039 --> 00:08:46.710 align:middle line:84% called scikit-learn, and I think they might be a little bit big. 00:08:46.710 --> 00:08:47.850 align:middle line:90% Might be-- whatever. 00:08:47.850 --> 00:08:53.300 align:middle line:84% So make sure to run all first, and that will get you 00:08:53.300 --> 00:08:56.660 align:middle line:90% in a pretty decent shape. 00:08:56.660 --> 00:08:58.700 align:middle line:90% I'm going to get started on-- 00:08:58.700 --> 00:09:02.990 align:middle line:84% what we're talking about today is something 00:09:02.990 --> 00:09:04.800 align:middle line:90% called feature extraction. 00:09:04.800 --> 00:09:08.480 align:middle line:84% And so feature extraction is a conversion of elements, 00:09:08.480 --> 00:09:14.820 align:middle line:84% in this case, musical elements, into single numbers or lists, 00:09:14.820 --> 00:09:18.110 align:middle line:84% arrays, vectors, tensors, whatever you want to call them, 00:09:18.110 --> 00:09:19.730 align:middle line:90% of numbers. 00:09:19.730 --> 00:09:23.130 align:middle line:84% It is the most boring thing I will teach in the class. 00:09:23.130 --> 00:09:25.877 align:middle line:84% And please don't correct me if you're like, Cuthbert, 00:09:25.877 --> 00:09:27.960 align:middle line:84% there's plenty of other boring things you've done. 00:09:27.960 --> 00:09:31.560 align:middle line:84% But no, it's really one of the most boring things in the world. 00:09:31.560 --> 00:09:34.550 align:middle line:90% And here's a reason to hate it. 00:09:34.550 --> 00:09:37.620 align:middle line:84% It really sucks the lifeblood out of music. 00:09:37.620 --> 00:09:39.530 align:middle line:84% Everything that you love about music 00:09:39.530 --> 00:09:42.690 align:middle line:84% is going to be converted into a few numbers. 00:09:42.690 --> 00:09:44.720 align:middle line:90% It is-- sucks the life out. 00:09:44.720 --> 00:09:47.130 align:middle line:90% It is the zombie of music world. 00:09:47.130 --> 00:09:48.270 align:middle line:90% Adam's very excited. 00:09:48.270 --> 00:09:49.850 align:middle line:90% Yes, zombie invasion! 00:09:49.850 --> 00:09:52.920 align:middle line:84% But there's a lot of amazing things, 00:09:52.920 --> 00:09:56.370 align:middle line:84% we know, happening in artificial intelligence, machine learning, 00:09:56.370 --> 00:09:58.770 align:middle line:84% deep learning, algorithms, all these things. 00:09:58.770 --> 00:10:03.120 align:middle line:84% And as far as I know, they all have one thing in common. 00:10:03.120 --> 00:10:06.210 align:middle line:84% And that is, they work on collections, 00:10:06.210 --> 00:10:09.740 align:middle line:90% big collections of numbers. 00:10:09.740 --> 00:10:12.710 align:middle line:84% There's some ways that you can work 00:10:12.710 --> 00:10:17.210 align:middle line:84% with images that, effectively, are kind of like two- 00:10:17.210 --> 00:10:18.960 align:middle line:84% or multi-dimensional arrays of numbers. 00:10:18.960 --> 00:10:21.030 align:middle line:84% But basically, they're all going to be numbers. 00:10:21.030 --> 00:10:29.210 align:middle line:84% So when I put some music notation on the board, 00:10:29.210 --> 00:10:31.110 align:middle line:84% we're not going to be able to work with that, 00:10:31.110 --> 00:10:32.910 align:middle line:84% even an image of it, even a JPEG, 00:10:32.910 --> 00:10:35.550 align:middle line:84% unless we convert it to pixels and convert it to numbers. 00:10:35.550 --> 00:10:38.090 align:middle line:84% When you hear a sound, [VOCALIZING] 00:10:38.090 --> 00:10:39.560 align:middle line:84% something like that, you're going 00:10:39.560 --> 00:10:45.060 align:middle line:84% to have to convert that into numbers in some ways. 00:10:45.060 --> 00:10:49.250 align:middle line:84% So we're going to be just working on converting things 00:10:49.250 --> 00:10:49.750 align:middle line:90% to numbers. 00:10:49.750 --> 00:10:54.470 align:middle line:90% So it's pretty simple, and-- 00:10:54.470 --> 00:10:58.650 align:middle line:84% but how to do it, and what to extract, 00:10:58.650 --> 00:11:00.270 align:middle line:84% and how to make your extractions, 00:11:00.270 --> 00:11:04.180 align:middle line:84% is the deep art that takes a while. 00:11:04.180 --> 00:11:08.070 align:middle line:84% Now, I'll get to a little bit later some of the things 00:11:08.070 --> 00:11:09.540 align:middle line:84% that people are doing today, where 00:11:09.540 --> 00:11:11.680 align:middle line:90% feature extraction is less-- 00:11:11.680 --> 00:11:14.710 align:middle line:84% or careful feature extraction is less important. 00:11:14.710 --> 00:11:18.160 align:middle line:84% And I'll tell you what those things tend to have in common, 00:11:18.160 --> 00:11:22.890 align:middle line:84% and that is, they work on much bigger data sets 00:11:22.890 --> 00:11:24.850 align:middle line:90% than what we have. 00:11:24.850 --> 00:11:27.850 align:middle line:84% So we're going to get started on this. 00:11:27.850 --> 00:11:29.890 align:middle line:90% Yep, that's my template here. 00:11:29.890 --> 00:11:32.980 align:middle line:84% And we're going to start with, why not? 00:11:32.980 --> 00:11:36.400 align:middle line:84% We always start with our favorite Bach chorale. 00:11:36.400 --> 00:11:37.590 align:middle line:90% What's the number? 00:11:37.590 --> 00:11:38.576 align:middle line:90% STUDENT: 66.6. 00:11:38.576 --> 00:11:40.800 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: 66.6. 00:11:40.800 --> 00:11:46.050 align:middle line:90% Conjure up the bwv66.6. 00:11:46.050 --> 00:11:47.915 align:middle line:84% I'm going to show it, get it back. 00:11:47.915 --> 00:11:59.640 align:middle line:90% 00:11:59.640 --> 00:12:04.440 align:middle line:90% OK, great. 00:12:04.440 --> 00:12:08.140 align:middle line:84% So we're going to do some manual feature extraction first. 00:12:08.140 --> 00:12:10.720 align:middle line:84% We're going to convert this into numbers. 00:12:10.720 --> 00:12:14.130 align:middle line:84% I'm going to extract two things from this, number of parts 00:12:14.130 --> 00:12:18.930 align:middle line:84% and initial number of beats per measure. 00:12:18.930 --> 00:12:22.243 align:middle line:90% So let's just do our-- 00:12:22.243 --> 00:12:24.160 align:middle line:84% You don't have to type this if you don't want. 00:12:24.160 --> 00:12:25.860 align:middle line:90% numParts equals 4. 00:12:25.860 --> 00:12:27.100 align:middle line:90% Time signature. 00:12:27.100 --> 00:12:29.140 align:middle line:90% Numerator equals 4. 00:12:29.140 --> 00:12:32.670 align:middle line:84% There, I have done manual feature extraction, 00:12:32.670 --> 00:12:36.900 align:middle line:84% and describing the piece as two numbers. 00:12:36.900 --> 00:12:40.740 align:middle line:84% They seem to look like 4-4, but I didn't ever just extract 00:12:40.740 --> 00:12:41.500 align:middle line:90% the denominator. 00:12:41.500 --> 00:12:43.625 align:middle line:84% That just happens to be coincidence they're both 4. 00:12:43.625 --> 00:12:47.860 align:middle line:84% So I don't know if we'd even call this feature extraction. 00:12:47.860 --> 00:12:49.810 align:middle line:90% But you can see, it's pretty-- 00:12:49.810 --> 00:12:52.990 align:middle line:84% we were not talking about very, very much here. 00:12:52.990 --> 00:12:59.110 align:middle line:84% So I'm going to define an automatic feature extractor. 00:12:59.110 --> 00:13:01.380 align:middle line:90% So we'll call it extract_meter. 00:13:01.380 --> 00:13:03.870 align:middle line:84% And extract_meter, you can save time 00:13:03.870 --> 00:13:05.560 align:middle line:84% by not putting these annotations. 00:13:05.560 --> 00:13:07.720 align:middle line:84% It takes in some sort of a stream. 00:13:07.720 --> 00:13:11.440 align:middle line:84% I'm calling it score, even though it's not. 00:13:11.440 --> 00:13:14.350 align:middle line:90% And it returns two integers. 00:13:14.350 --> 00:13:18.600 align:middle line:84% What do you think the two integers are going to be, 00:13:18.600 --> 00:13:21.108 align:middle line:90% if you're extracting the meter? 00:13:21.108 --> 00:13:21.900 align:middle line:90% STUDENT: Numerator. 00:13:21.900 --> 00:13:23.110 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: Numerator and denominator, 00:13:23.110 --> 00:13:23.610 align:middle line:90% great. 00:13:23.610 --> 00:13:24.120 align:middle line:90% Super. 00:13:24.120 --> 00:13:27.150 align:middle line:84% So we can say, all the meters in the piece 00:13:27.150 --> 00:13:32.970 align:middle line:84% equals the score meter, TimeSignature, recurse. 00:13:32.970 --> 00:13:38.790 align:middle line:84% And the first one, we'll just get the first one. first_meter, 00:13:38.790 --> 00:13:43.480 align:middle line:90% all_meters 0. 00:13:43.480 --> 00:13:48.850 align:middle line:84% And we'll return first_meter's numerator 00:13:48.850 --> 00:13:52.130 align:middle line:90% and first_meter's denominator. 00:13:52.130 --> 00:13:53.840 align:middle line:84% I know I don't need the parentheses, 00:13:53.840 --> 00:13:56.970 align:middle line:84% but I just kind of got used to that. 00:13:56.970 --> 00:13:59.590 align:middle line:90% 00:13:59.590 --> 00:14:04.930 align:middle line:84% So I'll let people get caught up. 00:14:04.930 --> 00:14:08.140 align:middle line:84% Now, by the way, how many time signatures 00:14:08.140 --> 00:14:10.300 align:middle line:90% does this little chorale have? 00:14:10.300 --> 00:14:13.110 align:middle line:84% Are we going overkill by getting the first one? 00:14:13.110 --> 00:14:17.830 align:middle line:90% 00:14:17.830 --> 00:14:19.510 align:middle line:90% Do you want to take a guess? 00:14:19.510 --> 00:14:21.365 align:middle line:90% You all know it by now. 00:14:21.365 --> 00:14:21.865 align:middle line:90% 1? 00:14:21.865 --> 00:14:24.550 align:middle line:90% 00:14:24.550 --> 00:14:27.770 align:middle line:84% It actually has 4, the way music 21 counts. 00:14:27.770 --> 00:14:29.420 align:middle line:90% 1, 2, 3, 4. 00:14:29.420 --> 00:14:30.410 align:middle line:90% They're all the same. 00:14:30.410 --> 00:14:33.070 align:middle line:84% But this is an error I make quite a bit 00:14:33.070 --> 00:14:36.170 align:middle line:84% often, that each of the parts has its own time signature. 00:14:36.170 --> 00:14:39.077 align:middle line:84% So this is something to be careful on. 00:14:39.077 --> 00:14:40.577 align:middle line:84% And now, we'll get to the first sort 00:14:40.577 --> 00:14:46.070 align:middle line:84% of ethical dilemma of feature extraction and machine learning, 00:14:46.070 --> 00:14:48.370 align:middle line:90% of which there will be many. 00:14:48.370 --> 00:14:50.230 align:middle line:90% And quite a number today. 00:14:50.230 --> 00:14:55.210 align:middle line:84% What happens if a score doesn't have any time signatures? 00:14:55.210 --> 00:14:56.330 align:middle line:90% Do we want this? 00:14:56.330 --> 00:14:59.950 align:middle line:90% Will this crash, first off? 00:14:59.950 --> 00:15:01.810 align:middle line:84% And if you think it will crash, tell 00:15:01.810 --> 00:15:06.895 align:middle line:84% me what the error that's probably going to be returned. 00:15:06.895 --> 00:15:08.270 align:middle line:84% STUDENT: List index out of range. 00:15:08.270 --> 00:15:08.890 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: List index 00:15:08.890 --> 00:15:10.400 align:middle line:90% out of range, or index error. 00:15:10.400 --> 00:15:12.500 align:middle line:90% Yep, that we're trying to get 0. 00:15:12.500 --> 00:15:20.270 align:middle line:84% So we can say something like, if not all_meters. 00:15:20.270 --> 00:15:22.700 align:middle line:84% Now, so here's the ethical dilemma. 00:15:22.700 --> 00:15:26.870 align:middle line:84% I'm going to return 0 comma 0 as a special value. 00:15:26.870 --> 00:15:31.190 align:middle line:84% That's also an integer, tuple of two integers, that says, 00:15:31.190 --> 00:15:35.260 align:middle line:84% I have no time signature because I've 00:15:35.260 --> 00:15:37.640 align:middle line:84% raised as a medieval musicologist, 00:15:37.640 --> 00:15:39.460 align:middle line:84% so I have Gregorian chant lying around, 00:15:39.460 --> 00:15:42.320 align:middle line:84% and other meterless time signature pieces. 00:15:42.320 --> 00:15:46.820 align:middle line:84% But the other reason is that we sometimes, 00:15:46.820 --> 00:15:50.630 align:middle line:84% in case of an error or an out-of-bound value, 00:15:50.630 --> 00:15:53.960 align:middle line:84% we return some sort of, we call it the sentinel, 00:15:53.960 --> 00:15:57.340 align:middle line:84% some kind of special meaning number that says 00:15:57.340 --> 00:15:59.740 align:middle line:90% some sort of an error happened. 00:15:59.740 --> 00:16:00.340 align:middle line:90% Why? 00:16:00.340 --> 00:16:04.635 align:middle line:84% Because you're going to someday have a gigantic data set. 00:16:04.635 --> 00:16:06.760 align:middle line:84% Maybe you're going to be doing this up in the cloud 00:16:06.760 --> 00:16:10.093 align:middle line:84% somewhere or something, and you're going to have-- 00:16:10.093 --> 00:16:13.060 align:middle line:90% 00:16:13.060 --> 00:16:15.680 align:middle line:84% you're going to be letting it run overnight, 00:16:15.680 --> 00:16:18.020 align:middle line:84% consuming a huge amount of processor money. 00:16:18.020 --> 00:16:21.460 align:middle line:84% And eventually, there's going to be some bad data in there that's 00:16:21.460 --> 00:16:23.500 align:middle line:90% going to crash your system. 00:16:23.500 --> 00:16:28.030 align:middle line:84% And you want to be like, OK, I just let this run for 47 hours, 00:16:28.030 --> 00:16:30.493 align:middle line:84% and now I have to start all over again. 00:16:30.493 --> 00:16:32.410 align:middle line:84% Because, of course, I didn't save intermediary 00:16:32.410 --> 00:16:35.610 align:middle line:84% values, because I'm a theoretical person, not 00:16:35.610 --> 00:16:37.120 align:middle line:90% a practical programmer. 00:16:37.120 --> 00:16:39.340 align:middle line:90% But always save. 00:16:39.340 --> 00:16:41.460 align:middle line:84% Write to disk every once in a while. 00:16:41.460 --> 00:16:45.840 align:middle line:84% But so this type of thing helps to save you 00:16:45.840 --> 00:16:47.440 align:middle line:90% from a little bit of bad data. 00:16:47.440 --> 00:16:51.780 align:middle line:84% What it also does, though, is it can swallow up systematic errors 00:16:51.780 --> 00:16:52.658 align:middle line:90% in your thought. 00:16:52.658 --> 00:16:53.700 align:middle line:90% You know what I'm saying? 00:16:53.700 --> 00:16:57.450 align:middle line:90% That you didn't realize that-- 00:16:57.450 --> 00:16:59.280 align:middle line:84% the one that I always get is, I'm 00:16:59.280 --> 00:17:01.170 align:middle line:84% looking at the pitch of everything 00:17:01.170 --> 00:17:03.310 align:middle line:84% in the piece, pitch of everything in there. 00:17:03.310 --> 00:17:07.540 align:middle line:84% And then I forget that these structures called chords exist, 00:17:07.540 --> 00:17:09.790 align:middle line:84% which don't just have a pitch, they have multiple. 00:17:09.790 --> 00:17:14.220 align:middle line:84% And so I could end up silently throwing out all of my data 00:17:14.220 --> 00:17:18.390 align:middle line:84% about chords, rather than realizing I have a mental model 00:17:18.390 --> 00:17:19.270 align:middle line:90% problem in there. 00:17:19.270 --> 00:17:21.630 align:middle line:84% And this is some of the things that 00:17:21.630 --> 00:17:25.560 align:middle line:84% can lead to AI that doesn't include all of us, 00:17:25.560 --> 00:17:28.050 align:middle line:84% that programmers forget that women exist 00:17:28.050 --> 00:17:29.650 align:middle line:90% or that non-white people exist. 00:17:29.650 --> 00:17:32.800 align:middle line:84% And therefore, oh, that's data that's out of bounds. 00:17:32.800 --> 00:17:35.540 align:middle line:84% So we'll go with a default response, 00:17:35.540 --> 00:17:39.560 align:middle line:84% rather than learning how to deal with facial recognition 00:17:39.560 --> 00:17:40.560 align:middle line:90% on things like that. 00:17:40.560 --> 00:17:44.000 align:middle line:90% So here, it's not such a big-- 00:17:44.000 --> 00:17:46.760 align:middle line:84% I don't think that we're going to have 00:17:46.760 --> 00:17:50.450 align:middle line:84% a problem with systematically discriminating against Gregorian 00:17:50.450 --> 00:17:51.120 align:middle line:90% chant. 00:17:51.120 --> 00:17:54.500 align:middle line:84% But it's these kinds of issues that 00:17:54.500 --> 00:17:58.700 align:middle line:84% lead to so many of the other ethical issues 00:17:58.700 --> 00:18:02.390 align:middle line:90% in artificial intelligence. 00:18:02.390 --> 00:18:04.110 align:middle line:84% With that aside, we've all caught up. 00:18:04.110 --> 00:18:06.320 align:middle line:90% Let's see, did I, somebody-- 00:18:06.320 --> 00:18:10.280 align:middle line:84% and everybody just save me time when I type something wrong. 00:18:10.280 --> 00:18:10.920 align:middle line:90% Check it. 00:18:10.920 --> 00:18:12.480 align:middle line:90% But let's see if that works. 00:18:12.480 --> 00:18:13.340 align:middle line:90% Whew. 00:18:13.340 --> 00:18:15.290 align:middle line:90% extract_meter(bach). 00:18:15.290 --> 00:18:22.230 align:middle line:84% We have our first, our first automatic feature extractor. 00:18:22.230 --> 00:18:25.456 align:middle line:84% Is this topic, so far, as boring as I promised? 00:18:25.456 --> 00:18:26.300 align:middle line:90% STUDENT: No. 00:18:26.300 --> 00:18:28.592 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: No OK, Somebody likes it. 00:18:28.592 --> 00:18:29.700 align:middle line:90% Yeah, it's pretty good. 00:18:29.700 --> 00:18:31.470 align:middle line:84% But let's go with something else. 00:18:31.470 --> 00:18:33.820 align:middle line:84% And we're going to start with another ethical dilemma. 00:18:33.820 --> 00:18:38.140 align:middle line:84% We're going to be working with Ryan's Mammoth Collection. 00:18:38.140 --> 00:18:42.030 align:middle line:84% And I think, if you type "blooming," 00:18:42.030 --> 00:18:44.800 align:middle line:84% you should get something that's not an error. 00:18:44.800 --> 00:18:45.737 align:middle line:90% Do you? 00:18:45.737 --> 00:18:47.820 align:middle line:84% Because you've already-- if you've run everything, 00:18:47.820 --> 00:18:54.180 align:middle line:84% you've already run corpus, parse, BloomingMeadowsJig. 00:18:54.180 --> 00:18:56.670 align:middle line:90% And let's see that. 00:18:56.670 --> 00:19:01.750 align:middle line:84% So apologies we didn't get to the jig composition yet. 00:19:01.750 --> 00:19:04.210 align:middle line:84% I do plan to pick that up next week, 00:19:04.210 --> 00:19:06.870 align:middle line:84% but I figure that there's some other things that people 00:19:06.870 --> 00:19:09.480 align:middle line:84% need for their final projects, more importantly 00:19:09.480 --> 00:19:11.550 align:middle line:90% than working with that. 00:19:11.550 --> 00:19:15.120 align:middle line:84% So you all see something like this, 00:19:15.120 --> 00:19:18.190 align:middle line:84% but we're going to be back on jigs again today for a bit. 00:19:18.190 --> 00:19:20.010 align:middle line:84% Now, Ryan's Mammoth Collection is 00:19:20.010 --> 00:19:22.060 align:middle line:84% a little bit of an ethical dilemma, 00:19:22.060 --> 00:19:23.440 align:middle line:90% and I'm trying to replace it. 00:19:23.440 --> 00:19:30.660 align:middle line:84% It is a collection of, I think, something like 1.079 fiddle 00:19:30.660 --> 00:19:34.020 align:middle line:90% tunes from the 19th century. 00:19:34.020 --> 00:19:37.200 align:middle line:84% And that has a bunch of jigs, hornpipes, 00:19:37.200 --> 00:19:40.950 align:middle line:84% things that if you take Joe Maurer's Sea 00:19:40.950 --> 00:19:45.250 align:middle line:84% Shanties and English Folk Music class and all those things, 00:19:45.250 --> 00:19:48.570 align:middle line:90% you can learn more about them. 00:19:48.570 --> 00:19:54.120 align:middle line:84% It also is a large collection of musical experiences of 19th 00:19:54.120 --> 00:20:00.390 align:middle line:84% century African-Americans that has been parodied and/or 00:20:00.390 --> 00:20:03.720 align:middle line:84% appropriated into white minstrel shows. 00:20:03.720 --> 00:20:10.770 align:middle line:84% So I've tried to remove things that are offensive, 00:20:10.770 --> 00:20:13.890 align:middle line:84% on first sight, from the collection, titles 00:20:13.890 --> 00:20:15.130 align:middle line:90% and things like that. 00:20:15.130 --> 00:20:19.080 align:middle line:84% But the deeper history, a lot of these songs, 00:20:19.080 --> 00:20:21.700 align:middle line:84% has to do with a lot of appropriation. 00:20:21.700 --> 00:20:24.990 align:middle line:84% So I'm hoping to be able to get a collection 00:20:24.990 --> 00:20:29.580 align:middle line:84% to replace it at some point, because it's not exactly 00:20:29.580 --> 00:20:31.360 align:middle line:84% everything that I want to do with it. 00:20:31.360 --> 00:20:34.890 align:middle line:84% But anyhow, so we have-- for going through, we have the one 00:20:34.890 --> 00:20:37.020 align:middle line:90% jig, Blooming Meadows jig. 00:20:37.020 --> 00:20:44.220 align:middle line:84% And I gave you just a contrast, a reel, R-E-E-L, 00:20:44.220 --> 00:20:47.798 align:middle line:84% and you should already have the reel, I think, already loaded. 00:20:47.798 --> 00:20:49.090 align:middle line:90% So you don't need to type this. 00:20:49.090 --> 00:20:51.750 align:middle line:84% It's called the "7th Regiment Reel." 00:20:51.750 --> 00:20:54.785 align:middle line:90% And hopefully, I got that. 00:20:54.785 --> 00:20:58.020 align:middle line:90% 00:20:58.020 --> 00:20:59.696 align:middle line:84% And it looks a little bit different. 00:20:59.696 --> 00:21:04.380 align:middle line:90% 00:21:04.380 --> 00:21:06.420 align:middle line:84% So what we're going to do today is 00:21:06.420 --> 00:21:09.210 align:middle line:84% try to build a system that can tell the difference 00:21:09.210 --> 00:21:11.250 align:middle line:90% between a jig and a reel. 00:21:11.250 --> 00:21:13.470 align:middle line:90% Or a jig and-- 00:21:13.470 --> 00:21:16.800 align:middle line:84% try to fit it into the remaining 52 minutes 00:21:16.800 --> 00:21:18.820 align:middle line:90% between a jig and a not-jig. 00:21:18.820 --> 00:21:23.500 align:middle line:84% The whole, all of music is divided into jigs and not jigs. 00:21:23.500 --> 00:21:26.050 align:middle line:90% It's great. 00:21:26.050 --> 00:21:29.790 align:middle line:90% So we we'll be looking-- 00:21:29.790 --> 00:21:31.583 align:middle line:84% how would you define the difference? 00:21:31.583 --> 00:21:33.000 align:middle line:84% You can look down at your own so I 00:21:33.000 --> 00:21:35.550 align:middle line:84% don't have to keep scrolling back and forth. 00:21:35.550 --> 00:21:39.180 align:middle line:84% Just, if you're somebody who didn't know, 00:21:39.180 --> 00:21:45.022 align:middle line:84% hadn't encountered a reel or a jig before, what is the-- 00:21:45.022 --> 00:21:48.850 align:middle line:84% what would you say is an obvious difference? 00:21:48.850 --> 00:21:49.350 align:middle line:90% Matthew? 00:21:49.350 --> 00:21:50.430 align:middle line:90% Do you-- 00:21:50.430 --> 00:21:51.545 align:middle line:90% STUDENT: 2/4 versus 6/8? 00:21:51.545 --> 00:21:52.660 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: 2/4 versus 6/8. 00:21:52.660 --> 00:21:53.370 align:middle line:90% Good. 00:21:53.370 --> 00:21:56.382 align:middle line:90% Another signature? 00:21:56.382 --> 00:21:58.090 align:middle line:84% STUDENT: It's the essential one, I think. 00:21:58.090 --> 00:21:59.010 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: What's that? 00:21:59.010 --> 00:22:01.210 align:middle line:84% STUDENT: It's the essential one, the time signature. 00:22:01.210 --> 00:22:02.650 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: The same one, time signature? 00:22:02.650 --> 00:22:03.150 align:middle line:90% Good. 00:22:03.150 --> 00:22:03.940 align:middle line:90% Yeah. 00:22:03.940 --> 00:22:06.870 align:middle line:84% STUDENT: So the [INAUDIBLE] seems more syncopated. 00:22:06.870 --> 00:22:08.110 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: Which one seems more syncopated? 00:22:08.110 --> 00:22:08.710 align:middle line:90% STUDENT: The jig. 00:22:08.710 --> 00:22:09.630 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: The jig, a little bit more 00:22:09.630 --> 00:22:10.270 align:middle line:90% syncopated? 00:22:10.270 --> 00:22:12.660 align:middle line:84% Especially things-- there's always 00:22:12.660 --> 00:22:16.560 align:middle line:84% things like this that are not, technically, rhythmically 00:22:16.560 --> 00:22:19.270 align:middle line:84% syncopated, but kind of feel a little bit. 00:22:19.270 --> 00:22:21.490 align:middle line:90% Mm-hmm. 00:22:21.490 --> 00:22:21.990 align:middle line:90% Good. 00:22:21.990 --> 00:22:23.430 align:middle line:90% Anything else? 00:22:23.430 --> 00:22:24.363 align:middle line:90% Yeah, Marlon. 00:22:24.363 --> 00:22:25.155 align:middle line:90% [NOTIFICATION TONE] 00:22:25.155 --> 00:22:25.710 align:middle line:90% STUDENT: [INAUDIBLE] 00:22:25.710 --> 00:22:27.040 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: What's that? 00:22:27.040 --> 00:22:28.390 align:middle line:84% STUDENT: I said 16th versus eighth notes. 00:22:28.390 --> 00:22:29.930 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: 16th versus eighth notes. 00:22:29.930 --> 00:22:30.440 align:middle line:90% Thank you. 00:22:30.440 --> 00:22:32.023 align:middle line:84% I'm going to put that in, and I'm also 00:22:32.023 --> 00:22:37.030 align:middle line:84% going to disable my, my what do you call it, my notifications. 00:22:37.030 --> 00:22:40.150 align:middle line:84% Supposed to happen automatically. 00:22:40.150 --> 00:22:41.050 align:middle line:90% OK, yeah. 00:22:41.050 --> 00:22:42.410 align:middle line:90% 16th versus eighth notes. 00:22:42.410 --> 00:22:46.730 align:middle line:84% We're actually going to possibly use that as something. 00:22:46.730 --> 00:22:49.960 align:middle line:84% So we're going to try to have a system that 00:22:49.960 --> 00:22:51.580 align:middle line:90% can do the difference. 00:22:51.580 --> 00:22:55.480 align:middle line:84% You should have-- if you type ryan, 00:22:55.480 --> 00:22:56.930 align:middle line:90% you should be getting something. 00:22:56.930 --> 00:23:00.340 align:middle line:84% I'm going to really quickly just cut and paste 00:23:00.340 --> 00:23:03.560 align:middle line:84% what I did to get to ryan and explain it. 00:23:03.560 --> 00:23:05.990 align:middle line:84% I first created something called ryanM, 00:23:05.990 --> 00:23:09.410 align:middle line:84% where I find everything in Ryan's Mammoth Collection. 00:23:09.410 --> 00:23:13.990 align:middle line:84% And then what I've done is, here's 00:23:13.990 --> 00:23:16.640 align:middle line:84% another cut and paste, 1059, I think, 00:23:16.640 --> 00:23:20.830 align:middle line:84% yep, is that I've gone through everything in there. 00:23:20.830 --> 00:23:22.720 align:middle line:84% And said, first off, I only want to look 00:23:22.720 --> 00:23:24.340 align:middle line:90% at the first 500 of them. 00:23:24.340 --> 00:23:29.590 align:middle line:84% And then, I'm looking, without parsing, if the name of the file 00:23:29.590 --> 00:23:33.490 align:middle line:90% has jig in it. 00:23:33.490 --> 00:23:36.830 align:middle line:84% If it doesn't have jig in it, and if you're over file 100, 00:23:36.830 --> 00:23:37.400 align:middle line:90% continue. 00:23:37.400 --> 00:23:39.730 align:middle line:90% So why did I do all this stuff? 00:23:39.730 --> 00:23:44.360 align:middle line:84% Just so that I can get something that is approximately 50/50 jig, 00:23:44.360 --> 00:23:45.850 align:middle line:90% not jig. 00:23:45.850 --> 00:23:49.090 align:middle line:90% So we end up with about 200. 00:23:49.090 --> 00:23:54.680 align:middle line:84% Here, we can-- we end up with 212 pieces, 00:23:54.680 --> 00:23:58.490 align:middle line:84% and I happen to know that 59% of them are jigs. 00:23:58.490 --> 00:24:00.620 align:middle line:90% So that's pretty good. 00:24:00.620 --> 00:24:03.860 align:middle line:84% These are all what we call metadata entries, 00:24:03.860 --> 00:24:06.730 align:middle line:90% so they haven't been parsed yet. 00:24:06.730 --> 00:24:08.630 align:middle line:90% OK, great. 00:24:08.630 --> 00:24:10.630 align:middle line:90% Everybody with me? 00:24:10.630 --> 00:24:11.740 align:middle line:90% There. 00:24:11.740 --> 00:24:12.430 align:middle line:90% Great. 00:24:12.430 --> 00:24:18.190 align:middle line:84% I'm going to define a helper function that is-- 00:24:18.190 --> 00:24:20.740 align:middle line:84% how many ethical dilemmas are we doing today? 00:24:20.740 --> 00:24:23.830 align:middle line:84% This one's ethical, but this one's sort of a research thing 00:24:23.830 --> 00:24:29.260 align:middle line:84% called is_jig, which determines if something 00:24:29.260 --> 00:24:35.110 align:middle line:84% is a jig by looking at the file name 00:24:35.110 --> 00:24:40.600 align:middle line:84% and lowercasing it, and seeing if the word jig is in that. 00:24:40.600 --> 00:24:47.200 align:middle line:84% I've done this with rags before, ragtime, a really great piano 00:24:47.200 --> 00:24:48.580 align:middle line:90% genre. 00:24:48.580 --> 00:24:50.720 align:middle line:84% Scott Joplin, the most famous person. 00:24:50.720 --> 00:24:53.960 align:middle line:84% And a lot of rags have "rag" in the title. 00:24:53.960 --> 00:24:56.210 align:middle line:90% "Maple Leaf Rag," and so on. 00:24:56.210 --> 00:25:00.278 align:middle line:84% But then, there's "The Entertainer," [VOCALIZING] 00:25:00.278 --> 00:25:00.778 align:middle line:90% 00:25:00.778 --> 00:25:03.140 align:middle line:84% --which doesn't have the word "rag" in the title. 00:25:03.140 --> 00:25:11.150 align:middle line:84% So we're possibly having what we call a bad ground truth. 00:25:11.150 --> 00:25:14.140 align:middle line:84% The ground truth is the definition 00:25:14.140 --> 00:25:17.510 align:middle line:90% that is correct beforehand. 00:25:17.510 --> 00:25:20.420 align:middle line:90% And so, is this song a jig? 00:25:20.420 --> 00:25:21.740 align:middle line:90% Is this not a jig? 00:25:21.740 --> 00:25:24.990 align:middle line:84% This is probably not the greatest ground truth there. 00:25:24.990 --> 00:25:27.170 align:middle line:84% The other thing that will come up a lot 00:25:27.170 --> 00:25:31.460 align:middle line:84% is that there is some pieces that will be 00:25:31.460 --> 00:25:34.170 align:middle line:90% very hard to tell if they are. 00:25:34.170 --> 00:25:36.620 align:middle line:84% And two different, intelligent people 00:25:36.620 --> 00:25:38.760 align:middle line:84% could disagree on something that's a jig. 00:25:38.760 --> 00:25:41.820 align:middle line:84% This comes up a lot in genre classification, 00:25:41.820 --> 00:25:45.110 align:middle line:84% where people can't really decide, is this country 00:25:45.110 --> 00:25:47.160 align:middle line:90% or is this rock? 00:25:47.160 --> 00:25:48.930 align:middle line:90% Is this funk? 00:25:48.930 --> 00:25:50.490 align:middle line:90% Is this hip-hop? 00:25:50.490 --> 00:25:52.080 align:middle line:90% Is this romantic? 00:25:52.080 --> 00:25:54.890 align:middle line:90% Is this classical? 00:25:54.890 --> 00:25:58.850 align:middle line:84% So somebody once told me, when they 00:25:58.850 --> 00:26:02.340 align:middle line:84% realized I came from a music history training, musicology, 00:26:02.340 --> 00:26:04.790 align:middle line:84% they said, the worst part about musicologists getting 00:26:04.790 --> 00:26:08.810 align:middle line:84% into this field is that they give us really crappy ground 00:26:08.810 --> 00:26:10.105 align:middle line:90% truths. 00:26:10.105 --> 00:26:12.900 align:middle line:84% You all won't say what the right answer is. 00:26:12.900 --> 00:26:16.950 align:middle line:84% And that's, I hope, an ambiguity we like as humans. 00:26:16.950 --> 00:26:21.830 align:middle line:84% But anyhow, so we'll have our is_jig. 00:26:21.830 --> 00:26:26.730 align:middle line:84% We'll actually see, is_jig(blooming). 00:26:26.730 --> 00:26:28.830 align:middle line:84% I think that was the one that is a jig. 00:26:28.830 --> 00:26:30.420 align:middle line:90% And what was the other one? 00:26:30.420 --> 00:26:32.360 align:middle line:90% A reel. 00:26:32.360 --> 00:26:35.120 align:middle line:90% OK. 00:26:35.120 --> 00:26:41.480 align:middle line:84% So with our garbage in, garbage out, terrible ground truth, we-- 00:26:41.480 --> 00:26:43.380 align:middle line:90% seems to be working OK. 00:26:43.380 --> 00:26:47.340 align:middle line:84% And I happen to know, 127 of the things in there is jigs. 00:26:47.340 --> 00:26:49.650 align:middle line:84% So let's get out second feature extractor. 00:26:49.650 --> 00:26:51.240 align:middle line:84% This one's going to be pretty easy. 00:26:51.240 --> 00:26:53.940 align:middle line:84% We'll just get the number of sharps in the piece. 00:26:53.940 --> 00:26:57.380 align:middle line:84% We started working on jigs in G, so we 00:26:57.380 --> 00:27:00.300 align:middle line:84% happen to know that there are a whole bunch of them. 00:27:00.300 --> 00:27:01.500 align:middle line:90% So maybe we'll do this. 00:27:01.500 --> 00:27:08.050 align:middle line:84% And here, we can just do, again, a try, return the scores 00:27:08.050 --> 00:27:12.890 align:middle line:84% key.KeySignature, the number of sharps. 00:27:12.890 --> 00:27:15.177 align:middle line:84% If it's two flats, what's it going to return? 00:27:15.177 --> 00:27:16.010 align:middle line:90% STUDENT: Negative 2. 00:27:16.010 --> 00:27:17.420 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: Negative 2, yep. 00:27:17.420 --> 00:27:18.210 align:middle line:90% Great. 00:27:18.210 --> 00:27:21.140 align:middle line:84% And if there's no key signature-- 00:27:21.140 --> 00:27:23.820 align:middle line:84% this one's actually not a terrible assumption, right? 00:27:23.820 --> 00:27:25.670 align:middle line:84% Something without a key signature 00:27:25.670 --> 00:27:29.197 align:middle line:84% probably has a key signature of no sharps or flats. 00:27:29.197 --> 00:27:30.030 align:middle line:90% I don't know, maybe. 00:27:30.030 --> 00:27:32.600 align:middle line:84% Maybe we should, in this case, analyze the key 00:27:32.600 --> 00:27:34.770 align:middle line:90% and do something like that. 00:27:34.770 --> 00:27:37.140 align:middle line:84% But good enough for just a little bit. 00:27:37.140 --> 00:27:38.930 align:middle line:90% So let's see. 00:27:38.930 --> 00:27:43.400 align:middle line:84% The blooming, hopefully this will return 1. 00:27:43.400 --> 00:27:45.710 align:middle line:90% Oops. 00:27:45.710 --> 00:27:47.610 align:middle line:90% Oh, I know what I did. 00:27:47.610 --> 00:27:50.600 align:middle line:84% This is why you should never do something like this. 00:27:50.600 --> 00:27:56.970 align:middle line:84% Except IndexError, we said, and we didn't get an index error. 00:27:56.970 --> 00:27:57.990 align:middle line:90% What did we get? 00:27:57.990 --> 00:27:59.640 align:middle line:90% We got an AttributeError? 00:27:59.640 --> 00:28:01.520 align:middle line:84% Yep, there is no sharps, because I 00:28:01.520 --> 00:28:05.850 align:middle line:84% forgot to say that we want the first key signatures sharps. 00:28:05.850 --> 00:28:09.650 align:middle line:84% You can also do the dot 0 or bracket 0, 00:28:09.650 --> 00:28:10.860 align:middle line:90% however you'd like that. 00:28:10.860 --> 00:28:14.570 align:middle line:84% And so let me try the get_sharps again. 00:28:14.570 --> 00:28:15.690 align:middle line:90% Getting 1. 00:28:15.690 --> 00:28:18.700 align:middle line:90% And what's the answer for Bach? 00:28:18.700 --> 00:28:19.200 align:middle line:90% STUDENT: 3. 00:28:19.200 --> 00:28:20.620 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: 3, good. 00:28:20.620 --> 00:28:22.830 align:middle line:84% You can either do that by hand or make sure [INAUDIBLE]. 00:28:22.830 --> 00:28:23.330 align:middle line:90% Great. 00:28:23.330 --> 00:28:26.280 align:middle line:84% And we'll define one more feature extractor 00:28:26.280 --> 00:28:28.380 align:middle line:90% so that we can-- 00:28:28.380 --> 00:28:30.510 align:middle line:90% oops. 00:28:30.510 --> 00:28:33.000 align:middle line:84% We're actually making good time this-- 00:28:33.000 --> 00:28:38.370 align:middle line:84% we'll do eighth_fraction based on the observation 00:28:38.370 --> 00:28:42.460 align:middle line:84% that there's more eighth notes in one piece than the other. 00:28:42.460 --> 00:28:45.370 align:middle line:90% So we'll take in a stream.Score. 00:28:45.370 --> 00:28:47.170 align:middle line:84% And this time, we'll return a float. 00:28:47.170 --> 00:28:50.680 align:middle line:84% So usually, you can return an int, 00:28:50.680 --> 00:28:53.310 align:middle line:84% which we call a continuous-- a discrete vari-- 00:28:53.310 --> 00:28:56.740 align:middle line:84% answer or a float, which is a continuous variable. 00:28:56.740 --> 00:28:58.150 align:middle line:90% So let's just do this. 00:28:58.150 --> 00:28:59.983 align:middle line:84% We'll get the number of notes in the piece. 00:28:59.983 --> 00:29:00.900 align:middle line:90% Actually, you do this. 00:29:00.900 --> 00:29:03.900 align:middle line:84% You give me-- go ahead and try coding this for a little bit, 00:29:03.900 --> 00:29:06.280 align:middle line:84% because we are a couple minutes ahead of time. 00:29:06.280 --> 00:29:09.708 align:middle line:90% How do we get the returns? 00:29:09.708 --> 00:29:11.250 align:middle line:84% The fra-- you don't need to put this. 00:29:11.250 --> 00:29:18.530 align:middle line:84% Returns the fraction of notes in a score that are eighth notes. 00:29:18.530 --> 00:29:23.100 align:middle line:90% 00:29:23.100 --> 00:29:25.410 align:middle line:90% OK, you can keep doing yours. 00:29:25.410 --> 00:29:28.770 align:middle line:84% If you're stuck, I'm going to give what I do. 00:29:28.770 --> 00:29:29.720 align:middle line:90% Sign like-- 00:29:29.720 --> 00:29:32.084 align:middle line:90% [TYPING] 00:29:32.084 --> 00:29:46.470 align:middle line:90% 00:29:46.470 --> 00:29:49.020 align:middle line:84% Not the most efficient use of iterators. 00:29:49.020 --> 00:29:49.865 align:middle line:90% I've done two. 00:29:49.865 --> 00:29:52.300 align:middle line:90% [TYPING] 00:29:52.300 --> 00:30:09.130 align:middle line:90% 00:30:09.130 --> 00:30:10.830 align:middle line:90% STUDENT: You're missing an H. 00:30:10.830 --> 00:30:11.700 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: I'm-- 00:30:11.700 --> 00:30:12.520 align:middle line:90% where am I missing it? 00:30:12.520 --> 00:30:13.920 align:middle line:84% It's somewhere in eighths, right? 00:30:13.920 --> 00:30:14.773 align:middle line:90% STUDENT: Yeah. 00:30:14.773 --> 00:30:15.196 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: Where's-- 00:30:15.196 --> 00:30:16.042 align:middle line:90% STUDENT: So second to last. 00:30:16.042 --> 00:30:17.295 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: Ah, eighth. 00:30:17.295 --> 00:30:18.630 align:middle line:90% That's-- yep, eighths. 00:30:18.630 --> 00:30:20.400 align:middle line:90% Thank you. 00:30:20.400 --> 00:30:21.365 align:middle line:90% Appreciate that. 00:30:21.365 --> 00:30:30.060 align:middle line:90% 00:30:30.060 --> 00:30:35.940 align:middle line:84% And go ahead and run your num eighths on Blooming, on Reel, 00:30:35.940 --> 00:30:36.990 align:middle line:90% and on Bach. 00:30:36.990 --> 00:30:41.340 align:middle line:84% And use your Spidey sense to figure out 00:30:41.340 --> 00:30:46.090 align:middle line:84% if your numbers are pretty meaningful. 00:30:46.090 --> 00:30:50.070 align:middle line:84% If there's anything I can get from this class 00:30:50.070 --> 00:30:52.440 align:middle line:84% to just use in all of your programming things is, 00:30:52.440 --> 00:30:58.470 align:middle line:84% always use your intuition to see if something 00:30:58.470 --> 00:31:04.510 align:middle line:84% like that, something is working right. 00:31:04.510 --> 00:31:05.620 align:middle line:90% Ooh, wow. 00:31:05.620 --> 00:31:08.130 align:middle line:90% Look at that. 00:31:08.130 --> 00:31:10.920 align:middle line:90% The reel is incredibly low. 00:31:10.920 --> 00:31:12.880 align:middle line:90% Oops. 00:31:12.880 --> 00:31:14.990 align:middle line:84% Now I have too many eighth fracti-- 00:31:14.990 --> 00:31:17.395 align:middle line:84% I should have done it eighths fraction, but whatever. 00:31:17.395 --> 00:31:17.895 align:middle line:90% I'm-- 00:31:17.895 --> 00:31:22.360 align:middle line:90% 00:31:22.360 --> 00:31:24.700 align:middle line:84% That actually feels too low for that reel. 00:31:24.700 --> 00:31:25.910 align:middle line:90% Only 8%. 00:31:25.910 --> 00:31:30.190 align:middle line:84% But I guess there's a lot of 16th notes. 00:31:30.190 --> 00:31:33.180 align:middle line:84% OK, give me a snap if you're ready to move on. 00:31:33.180 --> 00:31:33.880 align:middle line:90% [SNAPPING] 00:31:33.880 --> 00:31:36.350 align:middle line:90% OK, hearing some snaps, not all. 00:31:36.350 --> 00:31:38.730 align:middle line:84% We'll give another 10 seconds or so. 00:31:38.730 --> 00:31:42.304 align:middle line:90% 00:31:42.304 --> 00:31:48.160 align:middle line:84% So so far, we have the first one that seems really 00:31:48.160 --> 00:31:50.990 align:middle line:84% to give, at least in this one piece, 00:31:50.990 --> 00:31:57.140 align:middle line:84% a big distinction among the three types of things. 00:31:57.140 --> 00:32:01.090 align:middle line:90% 00:32:01.090 --> 00:32:06.070 align:middle line:84% OK, now I'm going to wade into something that people don't 00:32:06.070 --> 00:32:08.950 align:middle line:84% always say, is this a feature extractor 00:32:08.950 --> 00:32:10.760 align:middle line:84% because it returns a number or something? 00:32:10.760 --> 00:32:12.800 align:middle line:84% Or is, when you put them all together, 00:32:12.800 --> 00:32:14.390 align:middle line:90% is that the feature extractor? 00:32:14.390 --> 00:32:17.205 align:middle line:84% I'm not sure, but I'll use this term. 00:32:17.205 --> 00:32:18.580 align:middle line:84% So the feature extractor is going 00:32:18.580 --> 00:32:22.070 align:middle line:84% to run all of our different little feature extractors. 00:32:22.070 --> 00:32:24.700 align:middle line:84% And in this case, I'm going to say, 00:32:24.700 --> 00:32:27.090 align:middle line:84% it's going to return a tuple of-- 00:32:27.090 --> 00:32:29.210 align:middle line:84% and I'll tell you what these are in a bit. 00:32:29.210 --> 00:32:31.120 align:middle line:84% But here, I'll write this out here. 00:32:31.120 --> 00:32:37.320 align:middle line:84% A string, two ints, a float, an int, and an int. 00:32:37.320 --> 00:32:40.720 align:middle line:90% 00:32:40.720 --> 00:32:43.210 align:middle line:90% And what are these going to be? 00:32:43.210 --> 00:32:46.690 align:middle line:84% They're going to be the file name, which 00:32:46.690 --> 00:32:50.270 align:middle line:84% is pure metadata, just so that we can look at things later. 00:32:50.270 --> 00:32:53.620 align:middle line:84% They're going to be the numerator and the denominator 00:32:53.620 --> 00:32:58.238 align:middle line:84% of the piece using our function we wrote, extract meter. 00:32:58.238 --> 00:33:00.280 align:middle line:84% They're going to be-- what's the only thing we've 00:33:00.280 --> 00:33:02.523 align:middle line:84% written that returns a float as the answer? 00:33:02.523 --> 00:33:03.190 align:middle line:90% STUDENT: Eighth. 00:33:03.190 --> 00:33:04.070 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: What's that? 00:33:04.070 --> 00:33:04.660 align:middle line:90% STUDENT: Proportion of eighth. 00:33:04.660 --> 00:33:05.980 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: Eighth, yeah, eighth fraction, 00:33:05.980 --> 00:33:08.470 align:middle line:84% eighth proportion, whatever we are going to call it, 00:33:08.470 --> 00:33:12.970 align:middle line:84% the number of sharps, and then that last one we'll get to. 00:33:12.970 --> 00:33:16.510 align:middle line:84% That's going to be the ground truth. 00:33:16.510 --> 00:33:19.420 align:middle line:84% And it's doing too much JSON there. 00:33:19.420 --> 00:33:19.930 align:middle line:90% Stop 00:33:19.930 --> 00:33:21.860 align:middle line:90% Doing trailing commas. 00:33:21.860 --> 00:33:26.170 align:middle line:84% So the file name, you get this from a score's metadata 00:33:26.170 --> 00:33:28.460 align:middle line:84% object, which we haven't really talked about, 00:33:28.460 --> 00:33:32.420 align:middle line:90% it's file path split. 00:33:32.420 --> 00:33:33.880 align:middle line:84% I'm just going to get the last bit 00:33:33.880 --> 00:33:37.500 align:middle line:84% so I'm not showing my entire directory tree to everything. 00:33:37.500 --> 00:33:40.090 align:middle line:84% So I'll just get the last thing after the slash. 00:33:40.090 --> 00:33:42.980 align:middle line:84% Those of you who know Python's path library, 00:33:42.980 --> 00:33:46.540 align:middle line:90% you can probably do better. 00:33:46.540 --> 00:33:50.170 align:middle line:84% The numerator and the denominator 00:33:50.170 --> 00:33:56.650 align:middle line:84% will extract meter from the score. 00:33:56.650 --> 00:34:04.120 align:middle line:84% The eighths we'll say is the eighth fraction of the score. 00:34:04.120 --> 00:34:07.580 align:middle line:90% Sharps feature is get sharps. 00:34:07.580 --> 00:34:09.239 align:middle line:84% You can call them whatever you'd like. 00:34:09.239 --> 00:34:17.090 align:middle line:84% And then the ground truth, is it a jig or is it not? 00:34:17.090 --> 00:34:19.290 align:middle line:90% But it has to be a number. 00:34:19.290 --> 00:34:24.469 align:middle line:90% So we'll say integer of is jig. 00:34:24.469 --> 00:34:32.480 align:middle line:84% You could also say, if you like this Python, 1 if jig, else 0. 00:34:32.480 --> 00:34:34.520 align:middle line:84% Sometimes explicit's a little bit nicer 00:34:34.520 --> 00:34:36.449 align:middle line:90% to see what that's going to be. 00:34:36.449 --> 00:34:40.949 align:middle line:84% And then we want to return all that file name numerator, 00:34:40.949 --> 00:34:44.630 align:middle line:90% denominator, eighths-- 00:34:44.630 --> 00:34:45.929 align:middle line:90% I should do quarters. 00:34:45.929 --> 00:34:47.090 align:middle line:90% I can spell quarter-- 00:34:47.090 --> 00:34:50.400 align:middle line:90% feature, and ground truth. 00:34:50.400 --> 00:34:55.110 align:middle line:84% I'll let everybody get caught up while I debug my own system. 00:34:55.110 --> 00:34:58.740 align:middle line:84% Whoa, some fast typists in this class. 00:34:58.740 --> 00:35:01.440 align:middle line:90% Feature extractor on bach. 00:35:01.440 --> 00:35:03.110 align:middle line:90% Let me just see-- 00:35:03.110 --> 00:35:09.125 align:middle line:90% feature extractor on blooming. 00:35:09.125 --> 00:35:09.625 align:middle line:90% OK. 00:35:09.625 --> 00:35:15.830 align:middle line:90% 00:35:15.830 --> 00:35:17.337 align:middle line:90% ABC file extension, yep. 00:35:17.337 --> 00:35:18.170 align:middle line:90% STUDENT: [INAUDIBLE] 00:35:18.170 --> 00:35:19.628 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: Yeah. 00:35:19.628 --> 00:35:22.140 align:middle line:84% A lot of these are all written as ABC files. 00:35:22.140 --> 00:35:24.500 align:middle line:84% We didn't really cover-- no, we didn't cover ABC files 00:35:24.500 --> 00:35:29.480 align:middle line:84% because they're tiny bit more complex than the easiest 00:35:29.480 --> 00:35:33.740 align:middle line:84% to encode handwritten things but not complex 00:35:33.740 --> 00:35:38.240 align:middle line:84% enough as like musicxml or something. 00:35:38.240 --> 00:35:41.850 align:middle line:84% OK, so blooming metals jig, it's in 6/8. 00:35:41.850 --> 00:35:45.930 align:middle line:84% 77% of its notes are eighth notes. 00:35:45.930 --> 00:35:46.880 align:middle line:90% It's in one sharp. 00:35:46.880 --> 00:35:50.366 align:middle line:84% And that last one indicates what again? 00:35:50.366 --> 00:35:51.320 align:middle line:90% STUDENT: Jig. 00:35:51.320 --> 00:35:54.170 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: Yep, so the one indicates jig, 00:35:54.170 --> 00:35:54.720 align:middle line:90% right? 00:35:54.720 --> 00:35:55.370 align:middle line:90% Great. 00:35:55.370 --> 00:36:00.980 align:middle line:84% OK, so what I'm going to do right now 00:36:00.980 --> 00:36:03.870 align:middle line:84% is I'm just going to run this for a second. 00:36:03.870 --> 00:36:06.530 align:middle line:90% So how would I get everything? 00:36:06.530 --> 00:36:10.160 align:middle line:90% So we can do metadata. 00:36:10.160 --> 00:36:13.460 align:middle line:84% I'm just doing this so we can see how we do this. 00:36:13.460 --> 00:36:19.860 align:middle line:84% In the Ryan thing, piece equals metadata_entry.parse. 00:36:19.860 --> 00:36:26.750 align:middle line:84% And then we'll print the feature extractor on the piece. 00:36:26.750 --> 00:36:29.720 align:middle line:90% Let me just try to see. 00:36:29.720 --> 00:36:32.090 align:middle line:90% Yep. 00:36:32.090 --> 00:36:34.760 align:middle line:84% Because of the way I did it, you'll 00:36:34.760 --> 00:36:37.710 align:middle line:84% see that at first, there's not too many jigs. 00:36:37.710 --> 00:36:43.940 align:middle line:90% And then later on, there's jigs. 00:36:43.940 --> 00:36:50.940 align:middle line:84% OK, now, if you scroll back up to the top of your-- sorry, 00:36:50.940 --> 00:36:54.150 align:middle line:84% I'm going to have to get my blank template for a second. 00:36:54.150 --> 00:36:56.330 align:middle line:90% Oops, that's way smaller. 00:36:56.330 --> 00:36:59.700 align:middle line:84% There should be, on the top of your template 00:36:59.700 --> 00:37:02.540 align:middle line:84% if you downloaded it, something called write_ryan. 00:37:02.540 --> 00:37:06.390 align:middle line:84% And you just need to change your path 00:37:06.390 --> 00:37:11.080 align:middle line:84% to someplace that you're going to be able to remember it. 00:37:11.080 --> 00:37:14.880 align:middle line:84% So change that, unless you've created a Cuthbert 00:37:14.880 --> 00:37:17.670 align:middle line:84% user because it makes it easier for you to cut and paste 00:37:17.670 --> 00:37:18.825 align:middle line:90% my code. 00:37:18.825 --> 00:37:20.200 align:middle line:84% You probably want to change that. 00:37:20.200 --> 00:37:23.070 align:middle line:84% And then you can run this while I explain 00:37:23.070 --> 00:37:24.320 align:middle line:90% a little bit what it is. 00:37:24.320 --> 00:37:27.245 align:middle line:90% 00:37:27.245 --> 00:37:31.065 align:middle line:84% Is write_ryan working for people when you get that? 00:37:31.065 --> 00:37:31.565 align:middle line:90% OK. 00:37:31.565 --> 00:37:34.930 align:middle line:90% 00:37:34.930 --> 00:37:35.430 align:middle line:90% Great. 00:37:35.430 --> 00:37:38.310 align:middle line:90% 00:37:38.310 --> 00:37:45.185 align:middle line:84% So I'm going to have my own version here. 00:37:45.185 --> 00:37:55.140 align:middle line:90% 00:37:55.140 --> 00:37:56.227 align:middle line:90% STUDENT: [INAUDIBLE] 00:37:56.227 --> 00:37:58.560 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: The write_ryan( ) command. 00:37:58.560 --> 00:38:03.910 align:middle line:84% W-- you don't have to do any of this, but it'll be kind of fun 00:38:03.910 --> 00:38:13.665 align:middle line:84% because then you'll have write_ryan and start. 00:38:13.665 --> 00:38:16.300 align:middle line:90% 00:38:16.300 --> 00:38:18.940 align:middle line:84% OK, I've seen more people waiting 00:38:18.940 --> 00:38:22.430 align:middle line:84% than typing for a little bit, so I'll just explain what we do. 00:38:22.430 --> 00:38:24.710 align:middle line:84% Some of these things, if you work with files, 00:38:24.710 --> 00:38:27.460 align:middle line:84% you probably know them pretty well. 00:38:27.460 --> 00:38:29.960 align:middle line:84% This is really not necessarily the best way to do this, 00:38:29.960 --> 00:38:31.940 align:middle line:90% but I'm going to open two files. 00:38:31.940 --> 00:38:33.760 align:middle line:84% One of them is the file that we're 00:38:33.760 --> 00:38:37.040 align:middle line:84% going to use to train our classifier, 00:38:37.040 --> 00:38:40.300 align:middle line:84% the thing that is going to try to learn if something is a jig 00:38:40.300 --> 00:38:41.060 align:middle line:90% or not. 00:38:41.060 --> 00:38:44.080 align:middle line:84% The other one is the file we're going 00:38:44.080 --> 00:38:49.430 align:middle line:84% to do to try to test whether it can do it or not. 00:38:49.430 --> 00:38:52.840 align:middle line:84% We're going to hope that we don't 00:38:52.840 --> 00:38:55.190 align:middle line:90% put the same data into both. 00:38:55.190 --> 00:38:58.130 align:middle line:84% So we're going to try to put some data in the testing, 00:38:58.130 --> 00:38:59.283 align:middle line:90% some data in the training. 00:38:59.283 --> 00:39:00.700 align:middle line:84% There's a number of different ways 00:39:00.700 --> 00:39:03.130 align:middle line:90% you can do this type of work. 00:39:03.130 --> 00:39:08.030 align:middle line:84% Sometimes you put 90% in the training and 10% in the testing. 00:39:08.030 --> 00:39:10.900 align:middle line:84% Sometimes you put 50/50, which is what we're going to do. 00:39:10.900 --> 00:39:15.310 align:middle line:84% Some of it has to do with how much data you have. 00:39:15.310 --> 00:39:17.450 align:middle line:90% So what I'm going to do-- 00:39:17.450 --> 00:39:18.700 align:middle line:90% maybe it'll be easier to see-- 00:39:18.700 --> 00:39:21.610 align:middle line:84% I'm just going to separate-- .tab means everything's going 00:39:21.610 --> 00:39:22.970 align:middle line:90% to be separated by tabs. 00:39:22.970 --> 00:39:26.440 align:middle line:84% So I'm going to first put a header line that 00:39:26.440 --> 00:39:28.550 align:middle line:90% says the names of the columns. 00:39:28.550 --> 00:39:29.930 align:middle line:90% These are just for us to read. 00:39:29.930 --> 00:39:32.680 align:middle line:84% So we have a file name, numerator, denominator, eighths, 00:39:32.680 --> 00:39:34.760 align:middle line:90% sharp, and isjig. 00:39:34.760 --> 00:39:37.280 align:middle line:84% And so we're going to write that to the top of it 00:39:37.280 --> 00:39:41.410 align:middle line:84% because a lot of the classifiers that work on tab data 00:39:41.410 --> 00:39:43.760 align:middle line:84% expect this to be the first line. 00:39:43.760 --> 00:39:47.320 align:middle line:84% Now, we're going to hope that when running this system, 00:39:47.320 --> 00:39:49.610 align:middle line:84% that it doesn't peek at the file name. 00:39:49.610 --> 00:39:53.590 align:middle line:84% And we'll see because obviously, file name would give it away. 00:39:53.590 --> 00:39:57.020 align:middle line:84% And then so the descriptions, what kind of data is this? 00:39:57.020 --> 00:39:58.700 align:middle line:90% The file name's a string. 00:39:58.700 --> 00:40:00.190 align:middle line:84% Then we have discrete, as we said 00:40:00.190 --> 00:40:03.200 align:middle line:84% for an integer, discrete, continuous discrete, discrete. 00:40:03.200 --> 00:40:09.800 align:middle line:84% So we'll write that out again to each of the two files. 00:40:09.800 --> 00:40:13.060 align:middle line:84% And then we have third line, which 00:40:13.060 --> 00:40:15.350 align:middle line:84% is mostly blank for all these things. 00:40:15.350 --> 00:40:18.340 align:middle line:84% And the first one just says metadata. 00:40:18.340 --> 00:40:21.020 align:middle line:84% That is to say, don't peek at the file name-- 00:40:21.020 --> 00:40:24.110 align:middle line:84% do not use the file name to try to classify this-- 00:40:24.110 --> 00:40:27.860 align:middle line:84% then normal information and then the class, 00:40:27.860 --> 00:40:30.640 align:middle line:84% the thing that we're trying to figure out. 00:40:30.640 --> 00:40:34.960 align:middle line:84% And we really, really hope that the artificial intelligence 00:40:34.960 --> 00:40:39.260 align:middle line:84% thing does not peek at the class when it's trying to look. 00:40:39.260 --> 00:40:42.670 align:middle line:84% You can really quickly create an incredible AI thing 00:40:42.670 --> 00:40:44.057 align:middle line:90% if you would allow it to peek. 00:40:44.057 --> 00:40:45.140 align:middle line:90% And so we'll put that out. 00:40:45.140 --> 00:40:46.340 align:middle line:90% So that's just all headers. 00:40:46.340 --> 00:40:48.965 align:middle line:84% And that's why I didn't want to spend too much time writing it. 00:40:48.965 --> 00:40:50.540 align:middle line:90% Then what are we going to do? 00:40:50.540 --> 00:40:52.790 align:middle line:84% Same thing we were saying before, 00:40:52.790 --> 00:40:55.880 align:middle line:84% we're just going to get each piece into there. 00:40:55.880 --> 00:40:58.600 align:middle line:84% I've done enumerate for a reason so I can keep track 00:40:58.600 --> 00:41:00.670 align:middle line:90% of what the piece number is. 00:41:00.670 --> 00:41:06.770 align:middle line:84% And then we can extract the piece, all the features from it. 00:41:06.770 --> 00:41:09.820 align:middle line:84% And then we'll just write them out. 00:41:09.820 --> 00:41:11.990 align:middle line:84% In this case, we're going to write them as a string 00:41:11.990 --> 00:41:15.175 align:middle line:84% because we can only write strings out to a text file. 00:41:15.175 --> 00:41:19.660 align:middle line:84% Write each of the features out, tab separated. 00:41:19.660 --> 00:41:21.470 align:middle line:84% Actually, we'll get ready to write it. 00:41:21.470 --> 00:41:22.850 align:middle line:90% We'll just put it into a string. 00:41:22.850 --> 00:41:25.450 align:middle line:84% And then every other one, we're going 00:41:25.450 --> 00:41:28.160 align:middle line:84% to write to training and test, training and test. 00:41:28.160 --> 00:41:29.920 align:middle line:84% Probably, another way to do it would 00:41:29.920 --> 00:41:33.400 align:middle line:84% be to flip a coin each time and use a random number so you can 00:41:33.400 --> 00:41:35.330 align:middle line:90% see if things are going there. 00:41:35.330 --> 00:41:38.390 align:middle line:84% I'm not an expert on this type of work. 00:41:38.390 --> 00:41:41.750 align:middle line:90% But that's how that works. 00:41:41.750 --> 00:41:44.328 align:middle line:90% And I don't think-- 00:41:44.328 --> 00:41:46.120 align:middle line:84% Yeah, I don't think I'm going to run it now 00:41:46.120 --> 00:41:48.412 align:middle line:84% because I should have been running it during this time. 00:41:48.412 --> 00:41:49.070 align:middle line:90% You know what? 00:41:49.070 --> 00:41:53.770 align:middle line:84% I'm just going to put train data x and test data x 00:41:53.770 --> 00:41:59.720 align:middle line:84% so that while it's running, I have access to our baked one. 00:41:59.720 --> 00:42:04.820 align:middle line:84% So actually we can look at it, magic features. 00:42:04.820 --> 00:42:12.070 align:middle line:84% So we switch out of Python and look at the shell for a second-- 00:42:12.070 --> 00:42:15.520 align:middle line:84% train data-- wherever you put your thing. 00:42:15.520 --> 00:42:21.730 align:middle line:84% And I can see that I've written out-- 00:42:21.730 --> 00:42:24.065 align:middle line:84% it's kind of weird that the tabs don't show up. 00:42:24.065 --> 00:42:26.440 align:middle line:84% You would love to have something that would automatically 00:42:26.440 --> 00:42:26.940 align:middle line:90% put it out. 00:42:26.940 --> 00:42:29.920 align:middle line:84% But we have the file name, all our headers, description 00:42:29.920 --> 00:42:32.630 align:middle line:90% of the data, the metadata here. 00:42:32.630 --> 00:42:38.180 align:middle line:84% And then yeah, we have all of our information written out. 00:42:38.180 --> 00:42:40.160 align:middle line:90% So that's pretty good. 00:42:40.160 --> 00:42:42.910 align:middle line:90% Oh that did finish, great. 00:42:42.910 --> 00:42:45.370 align:middle line:90% OK. 00:42:45.370 --> 00:42:47.960 align:middle line:84% So now we've done our feature extraction. 00:42:47.960 --> 00:42:50.770 align:middle line:84% The feature extraction part of the lecture 00:42:50.770 --> 00:42:53.200 align:middle line:90% is done for a little bit. 00:42:53.200 --> 00:42:55.413 align:middle line:84% This is a small number of features. 00:42:55.413 --> 00:42:57.080 align:middle line:84% But I think that they're going to end up 00:42:57.080 --> 00:42:59.810 align:middle line:84% being halfway decent features because we 00:42:59.810 --> 00:43:01.620 align:middle line:90% thought about them a bit. 00:43:01.620 --> 00:43:04.340 align:middle line:84% And when the data is small, you want to think more 00:43:04.340 --> 00:43:05.190 align:middle line:90% about your features. 00:43:05.190 --> 00:43:08.040 align:middle line:84% So you installed something called Orange 3. 00:43:08.040 --> 00:43:09.830 align:middle line:90% But it's a little bit weird. 00:43:09.830 --> 00:43:12.210 align:middle line:90% It comes as capital Orange. 00:43:12.210 --> 00:43:13.410 align:middle line:90% I just I don't know why. 00:43:13.410 --> 00:43:15.630 align:middle line:90% I don't like things capital. 00:43:15.630 --> 00:43:19.610 align:middle line:84% So I always import orange as lowercase orange. 00:43:19.610 --> 00:43:21.530 align:middle line:90% Great. 00:43:21.530 --> 00:43:23.750 align:middle line:84% We're going to load our training data 00:43:23.750 --> 00:43:28.280 align:middle line:84% and our test data, so first training data, 00:43:28.280 --> 00:43:29.600 align:middle line:90% orange.data.table. 00:43:29.600 --> 00:43:32.840 align:middle line:90% 00:43:32.840 --> 00:43:35.630 align:middle line:84% A lot of these things are going to be very similar 00:43:35.630 --> 00:43:40.520 align:middle line:84% in a super modern artificial intelligence 00:43:40.520 --> 00:43:45.380 align:middle line:84% framework like PyTorch or TensorFlow 00:43:45.380 --> 00:43:46.340 align:middle line:90% or something like this. 00:43:46.340 --> 00:43:51.170 align:middle line:84% Similar idea, you want to load data in as a table. 00:43:51.170 --> 00:43:55.080 align:middle line:84% If it's too big to load on your all at once, 00:43:55.080 --> 00:43:58.170 align:middle line:84% then there's special techniques and stuff like that. 00:43:58.170 --> 00:44:00.630 align:middle line:84% But we'll load the training data in. 00:44:00.630 --> 00:44:04.480 align:middle line:84% And now when we look at it, it just looks a little bit nicer. 00:44:04.480 --> 00:44:08.150 align:middle line:90% 00:44:08.150 --> 00:44:11.570 align:middle line:90% Actually, where's our first-- 00:44:11.570 --> 00:44:13.250 align:middle line:90% I don't know-- 00:44:13.250 --> 00:44:16.170 align:middle line:90% 45 to 50? 00:44:16.170 --> 00:44:16.830 align:middle line:90% Does that work? 00:44:16.830 --> 00:44:17.570 align:middle line:90% Yeah. 00:44:17.570 --> 00:44:20.030 align:middle line:84% OK, so that might be a little bit better 00:44:20.030 --> 00:44:24.050 align:middle line:84% because we can see we have our features 00:44:24.050 --> 00:44:28.830 align:middle line:90% here two for 12% and 3 sharps. 00:44:28.830 --> 00:44:34.580 align:middle line:84% And this one happens to be a jig, so there. 00:44:34.580 --> 00:44:37.950 align:middle line:84% And so the afterward is what the class is. 00:44:37.950 --> 00:44:40.710 align:middle line:84% Great, and we'll do the exact same thing with the test data. 00:44:40.710 --> 00:44:42.450 align:middle line:84% And we'll call it variable test data. 00:44:42.450 --> 00:44:44.300 align:middle line:84% So I'm just going to copy and paste. 00:44:44.300 --> 00:44:46.070 align:middle line:84% And hopefully, the test data looks 00:44:46.070 --> 00:44:47.530 align:middle line:90% a lot like the training data. 00:44:47.530 --> 00:44:52.880 align:middle line:90% 00:44:52.880 --> 00:44:54.860 align:middle line:84% STUDENT: Should we change the file path? 00:44:54.860 --> 00:44:56.735 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: Oh, yeah, yeah. 00:44:56.735 --> 00:44:59.160 align:middle line:84% It looks a lot like the training data 00:44:59.160 --> 00:45:01.130 align:middle line:84% if we don't change the file path. 00:45:01.130 --> 00:45:02.460 align:middle line:90% You see how well it is. 00:45:02.460 --> 00:45:08.030 align:middle line:84% OK, so now we're going to create some, 00:45:08.030 --> 00:45:10.280 align:middle line:84% we'll call learners, classifiers, things 00:45:10.280 --> 00:45:14.040 align:middle line:90% that can classify the data. 00:45:14.040 --> 00:45:17.300 align:middle line:84% Here's the first one and I think, 00:45:17.300 --> 00:45:21.090 align:middle line:84% in some ways, the most important classifier that you'll ever use. 00:45:21.090 --> 00:45:24.120 align:middle line:84% And that is-- why did I switch into camelcase? 00:45:24.120 --> 00:45:26.060 align:middle line:84% OK, I guess I'm in camelcase now-- 00:45:26.060 --> 00:45:28.917 align:middle line:84% something called orange.classific 00:45:28.917 --> 00:45:30.275 align:middle line:90% ation.MajorityLearner. 00:45:30.275 --> 00:45:38.690 align:middle line:90% 00:45:38.690 --> 00:45:43.700 align:middle line:84% And what a majority learner is it's kind 00:45:43.700 --> 00:45:47.900 align:middle line:84% of the control or the placebo of your system 00:45:47.900 --> 00:45:51.270 align:middle line:84% to make sure that you don't get too excited by how 00:45:51.270 --> 00:45:53.650 align:middle line:90% good your system is working. 00:45:53.650 --> 00:45:57.040 align:middle line:84% What it does is it looks at the training data, 00:45:57.040 --> 00:45:59.020 align:middle line:90% which has hundreds-- 00:45:59.020 --> 00:46:02.108 align:middle line:84% what maybe the training data has-- 00:46:02.108 --> 00:46:02.650 align:middle line:90% I don't know. 00:46:02.650 --> 00:46:08.620 align:middle line:84% Let's say it has 51 jigs and 49 nonjigs. 00:46:08.620 --> 00:46:13.230 align:middle line:84% And it says, I will always say that any piece that's 00:46:13.230 --> 00:46:16.260 align:middle line:84% ever shown to me is a jig because then I will get over 00:46:16.260 --> 00:46:18.360 align:middle line:84% 50% right or I will get the highest 00:46:18.360 --> 00:46:20.530 align:middle line:90% answer of doing anything. 00:46:20.530 --> 00:46:24.630 align:middle line:84% So if you are not beating your majority learner, 00:46:24.630 --> 00:46:26.860 align:middle line:84% then you're not doing any classification. 00:46:26.860 --> 00:46:30.570 align:middle line:90% So the majority classifier-- 00:46:30.570 --> 00:46:33.430 align:middle line:84% you all who have taken a machine learning class, 00:46:33.430 --> 00:46:36.550 align:middle line:90% do you get taught this or no? 00:46:36.550 --> 00:46:37.050 align:middle line:90% Yeah. 00:46:37.050 --> 00:46:38.010 align:middle line:90% OK, good. 00:46:38.010 --> 00:46:43.030 align:middle line:84% Some people, this will be totally old story and not. 00:46:43.030 --> 00:46:47.140 align:middle line:84% So the majority learner is just a generic thing. 00:46:47.140 --> 00:46:48.870 align:middle line:84% And now we have a classifier, which 00:46:48.870 --> 00:46:51.130 align:middle line:90% is one trained on our data. 00:46:51.130 --> 00:46:56.490 align:middle line:84% So we're going to say that the majority classifier will now 00:46:56.490 --> 00:47:01.440 align:middle line:84% know how to work with this particular data. 00:47:01.440 --> 00:47:04.690 align:middle line:84% And then we'll use another one called k-nearest neighbors. 00:47:04.690 --> 00:47:08.470 align:middle line:84% This was-- I don't know-- maybe 20 years ago, 00:47:08.470 --> 00:47:13.743 align:middle line:84% the state of the art of machine learning. 00:47:13.743 --> 00:47:15.160 align:middle line:84% It's a little bit out of date now. 00:47:15.160 --> 00:47:18.240 align:middle line:84% But it does work pretty well on small data 00:47:18.240 --> 00:47:20.470 align:middle line:90% sets and few classifiers. 00:47:20.470 --> 00:47:23.700 align:middle line:84% So nothing that it's going to get you 00:47:23.700 --> 00:47:27.460 align:middle line:84% super excited at a top tech firm. 00:47:27.460 --> 00:47:30.910 align:middle line:90% But it still has its uses. 00:47:30.910 --> 00:47:33.990 align:middle line:84% So we're going to do the same exact same thing. 00:47:33.990 --> 00:47:36.030 align:middle line:84% Everything that I put from majority learner, 00:47:36.030 --> 00:47:39.060 align:middle line:84% and just put it as KNN neighbors. 00:47:39.060 --> 00:47:41.840 align:middle line:84% Look at the K, which might be 4, might be 3, 00:47:41.840 --> 00:47:44.190 align:middle line:84% might be 5, the nearest neighbors 00:47:44.190 --> 00:47:49.200 align:middle line:84% in a multi-dimensional vector space to your one piece 00:47:49.200 --> 00:47:55.050 align:middle line:84% and try to learn where in this space it goes. 00:47:55.050 --> 00:47:58.290 align:middle line:90% Snap if you're caught. 00:47:58.290 --> 00:48:01.620 align:middle line:84% Snap if you're not now, if you need a second. 00:48:01.620 --> 00:48:05.190 align:middle line:90% OK, good. 00:48:05.190 --> 00:48:05.790 align:middle line:90% Great. 00:48:05.790 --> 00:48:09.450 align:middle line:84% So I'm just going to keep track of how many correct 00:48:09.450 --> 00:48:11.860 align:middle line:84% the majority gets in a variable called majCorrect 00:48:11.860 --> 00:48:15.990 align:middle line:84% and how many correct the k-nearest neighbor gets 00:48:15.990 --> 00:48:18.100 align:middle line:90% in a variable called knnCorrect. 00:48:18.100 --> 00:48:24.240 align:middle line:84% And then total equals the length of the test data, not 00:48:24.240 --> 00:48:25.000 align:middle line:90% the training data. 00:48:25.000 --> 00:48:27.280 align:middle line:84% So now we're going to be working on this. 00:48:27.280 --> 00:48:31.480 align:middle line:84% And we should have something like this. 00:48:31.480 --> 00:48:38.490 align:middle line:84% By the way, I lost half an hour in preparing this lecture on, 00:48:38.490 --> 00:48:40.240 align:middle line:90% why do I keep getting over 100%? 00:48:40.240 --> 00:48:44.340 align:middle line:84% Always make sure to reclear out your number correct as you 00:48:44.340 --> 00:48:49.938 align:middle line:84% recode or something or make sure to do it within your loop. 00:48:49.938 --> 00:48:51.440 align:middle line:90% So good. 00:48:51.440 --> 00:48:56.030 align:middle line:84% So we have 106 things that we're going to be able to do now. 00:48:56.030 --> 00:48:59.500 align:middle line:90% That was for our loop. 00:48:59.500 --> 00:49:06.230 align:middle line:84% So we'll just do an enumeration of each row in the test data. 00:49:06.230 --> 00:49:09.930 align:middle line:84% So for i comma test row in enumerate test data. 00:49:09.930 --> 00:49:13.885 align:middle line:90% 00:49:13.885 --> 00:49:16.450 align:middle line:84% And then the majority learner is going 00:49:16.450 --> 00:49:18.950 align:middle line:90% to take a guess for each one. 00:49:18.950 --> 00:49:22.870 align:middle line:84% So majority's guessed is going to be the majority classifier 00:49:22.870 --> 00:49:26.320 align:middle line:84% of the test row, which it's always 00:49:26.320 --> 00:49:28.040 align:middle line:90% going to guess the same thing. 00:49:28.040 --> 00:49:29.470 align:middle line:84% I guess in this case, it's always 00:49:29.470 --> 00:49:31.900 align:middle line:84% going to guess jig because there's four more 00:49:31.900 --> 00:49:33.650 align:middle line:90% jigs than there are non jigs. 00:49:33.650 --> 00:49:38.050 align:middle line:84% And then our classifier, k-nearest neighbor classifier, 00:49:38.050 --> 00:49:43.630 align:middle line:84% will also take a guess, knnClassifier. 00:49:43.630 --> 00:49:45.550 align:middle line:84% I know somebody's probably saying "guess" 00:49:45.550 --> 00:49:48.440 align:middle line:84% is not the right word, but type works. 00:49:48.440 --> 00:49:55.310 align:middle line:84% And then we'll get the real is the testrows.get_class. 00:49:55.310 --> 00:50:00.200 align:middle line:84% So as you said, the class is either 1 or 0, 00:50:00.200 --> 00:50:01.760 align:middle line:90% something like that. 00:50:01.760 --> 00:50:06.200 align:middle line:84% And we can do print out something. 00:50:06.200 --> 00:50:10.130 align:middle line:90% But print-- how about this? 00:50:10.130 --> 00:50:19.180 align:middle line:84% We'll just say majGuess, knnGuess, and real. 00:50:19.180 --> 00:50:22.880 align:middle line:84% So I'll start this running for a second. 00:50:22.880 --> 00:50:24.940 align:middle line:84% But notice, we haven't put in anything 00:50:24.940 --> 00:50:26.600 align:middle line:84% on whether it got it right or not. 00:50:26.600 --> 00:50:28.930 align:middle line:90% But we can just look at that. 00:50:28.930 --> 00:50:32.590 align:middle line:90% Oh, one of them guesses. 00:50:32.590 --> 00:50:36.970 align:middle line:84% So you can see that the majority one for every single piece 00:50:36.970 --> 00:50:39.640 align:middle line:90% just puts out-- 00:50:39.640 --> 00:50:43.360 align:middle line:84% yep, it's always a jig, always a jig, always a jig. 00:50:43.360 --> 00:50:49.340 align:middle line:84% K-nearest neighbor says, I don't think this is a jig. 00:50:49.340 --> 00:50:50.690 align:middle line:90% And here's the correct answer-- 00:50:50.690 --> 00:50:52.460 align:middle line:90% not a jig, not a jig, not a jig. 00:50:52.460 --> 00:50:55.930 align:middle line:84% And here, k-nearest neighbor is learning that, ooh, this one 00:50:55.930 --> 00:50:57.260 align:middle line:90% looks like a jig. 00:50:57.260 --> 00:51:01.510 align:middle line:84% So I'll just modify this a tiny bit 00:51:01.510 --> 00:51:12.850 align:middle line:84% and say, if majGuess equals real, then majCorrect add 1. 00:51:12.850 --> 00:51:22.687 align:middle line:84% If knnGuess equals real, knnCorrect equals 1. 00:51:22.687 --> 00:51:24.770 align:middle line:84% When you're trying to figure out how to do things, 00:51:24.770 --> 00:51:26.210 align:middle line:90% you'll use more classifiers. 00:51:26.210 --> 00:51:27.200 align:middle line:90% There's a lot more. 00:51:27.200 --> 00:51:30.230 align:middle line:84% And so you probably want to do this in some of a loop. 00:51:30.230 --> 00:51:31.623 align:middle line:90% But great. 00:51:31.623 --> 00:51:37.595 align:middle line:90% 00:51:37.595 --> 00:51:38.095 align:middle line:90% Good. 00:51:38.095 --> 00:51:40.780 align:middle line:90% 00:51:40.780 --> 00:51:42.510 align:middle line:90% And so now we'll see how we did. 00:51:42.510 --> 00:51:43.720 align:middle line:90% How did majority do? 00:51:43.720 --> 00:51:48.230 align:middle line:90% 00:51:48.230 --> 00:51:49.940 align:middle line:90% Here, I'll do this. 00:51:49.940 --> 00:51:55.010 align:middle line:84% Majority majCorrect-- I always making 00:51:55.010 --> 00:51:56.605 align:middle line:90% things look nice for humans. 00:51:56.605 --> 00:52:05.930 align:middle line:90% 00:52:05.930 --> 00:52:09.620 align:middle line:84% Why am I anal enough to start already 00:52:09.620 --> 00:52:14.040 align:middle line:84% putting in some of a rounding thing to look? 00:52:14.040 --> 00:52:18.990 align:middle line:84% Because I do feel like when you see all those digits, 00:52:18.990 --> 00:52:23.610 align:middle line:84% it's really, really hard to figure out how important 00:52:23.610 --> 00:52:24.360 align:middle line:90% the difference is. 00:52:24.360 --> 00:52:26.270 align:middle line:84% So the majority, this is what I happen 00:52:26.270 --> 00:52:28.800 align:middle line:84% to know that there's 59% of them are jigs. 00:52:28.800 --> 00:52:32.490 align:middle line:84% And so the majority learner got 59% of them. 00:52:32.490 --> 00:52:37.640 align:middle line:84% And then we'll do the same thing for KNN. 00:52:37.640 --> 00:52:39.912 align:middle line:84% Here, I can cut and paste, right? 00:52:39.912 --> 00:52:41.670 align:middle line:90% We all know how to do that. 00:52:41.670 --> 00:52:43.450 align:middle line:90% What are we looking up? 00:52:43.450 --> 00:52:43.950 align:middle line:90% Great. 00:52:43.950 --> 00:52:47.100 align:middle line:90% 00:52:47.100 --> 00:52:49.750 align:middle line:84% And we want to make sure that that's knnCorrect. 00:52:49.750 --> 00:52:53.840 align:middle line:90% 00:52:53.840 --> 00:53:01.100 align:middle line:84% So our classifier is getting about 87% correct at this, just 00:53:01.100 --> 00:53:03.020 align:middle line:90% with these few little things. 00:53:03.020 --> 00:53:07.100 align:middle line:84% Now, this is always the exciting moment. 00:53:07.100 --> 00:53:12.090 align:middle line:84% And then the hard part is almost always getting it from 87 to 95. 00:53:12.090 --> 00:53:14.360 align:middle line:84% And the next harder part is getting it from 95 to 97 00:53:14.360 --> 00:53:18.110 align:middle line:84% and then going up and up and up from there. 00:53:18.110 --> 00:53:21.900 align:middle line:84% So that's the theoretical part aside. 00:53:21.900 --> 00:53:28.460 align:middle line:84% I'm going to really quickly just go through some things 00:53:28.460 --> 00:53:37.610 align:middle line:84% that music 21 has that just make some 00:53:37.610 --> 00:53:40.460 align:middle line:90% of this a little bit easier. 00:53:40.460 --> 00:53:44.040 align:middle line:84% So I know it's already imported since I imported star. 00:53:44.040 --> 00:53:46.050 align:middle line:84% But there's something called features. 00:53:46.050 --> 00:53:54.440 align:middle line:84% And what we can do is set equals a features.DataSet classLabel 00:53:54.440 --> 00:53:57.050 align:middle line:90% equals is_jig. 00:53:57.050 --> 00:54:00.650 align:middle line:84% And then we'll do the same thing with the test set. 00:54:00.650 --> 00:54:05.210 align:middle line:84% So far, it's basically the same type of thing. 00:54:05.210 --> 00:54:08.780 align:middle line:84% And then I believe if you type fes, 00:54:08.780 --> 00:54:10.970 align:middle line:90% you should have something there. 00:54:10.970 --> 00:54:12.470 align:middle line:90% I think I pasted this. 00:54:12.470 --> 00:54:15.210 align:middle line:90% 00:54:15.210 --> 00:54:17.480 align:middle line:84% These are built-in feature extractor. 00:54:17.480 --> 00:54:19.580 align:middle line:90% And each one of them has an ID. 00:54:19.580 --> 00:54:23.680 align:middle line:84% And no, I don't have these memorized, but something about-- 00:54:23.680 --> 00:54:27.050 align:middle line:90% 00:54:27.050 --> 00:54:29.477 align:middle line:90% I can't remember what these are. 00:54:29.477 --> 00:54:30.060 align:middle line:90% You know what? 00:54:30.060 --> 00:54:32.602 align:middle line:84% I'm just going to cut and paste because you already have this 00:54:32.602 --> 00:54:35.780 align:middle line:90% and put it here. 00:54:35.780 --> 00:54:37.800 align:middle line:90% STUDENT: You have fears? 00:54:37.800 --> 00:54:41.890 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: Fears, yes, that's my hidden-- 00:54:41.890 --> 00:54:42.970 align:middle line:90% thank you very much. 00:54:42.970 --> 00:54:44.140 align:middle line:90% Appreciate that. 00:54:44.140 --> 00:54:47.550 align:middle line:90% And think that should work. 00:54:47.550 --> 00:54:52.650 align:middle line:84% And so these are some features that have already been put in. 00:54:52.650 --> 00:54:56.680 align:middle line:84% So one of them, I don't know, what's the range of the piece? 00:54:56.680 --> 00:54:59.440 align:middle line:84% How far is it from the lowest note to the highest note? 00:54:59.440 --> 00:55:02.670 align:middle line:84% A lot of these, they're adapted from a really, really 00:55:02.670 --> 00:55:06.870 align:middle line:84% cool toolkit called jSymbolic part of jMIR 00:55:06.870 --> 00:55:12.630 align:middle line:84% by Cory Mckay up in Canada, and just ported them over 00:55:12.630 --> 00:55:13.930 align:middle line:90% because they're kind of neat. 00:55:13.930 --> 00:55:17.760 align:middle line:84% And then some of them are ones that I've written because I kind 00:55:17.760 --> 00:55:19.830 align:middle line:84% think that they're a little valuable 00:55:19.830 --> 00:55:22.360 align:middle line:84% or somebody in the Music 21 team has written. 00:55:22.360 --> 00:55:25.660 align:middle line:84% So here, we have a lot more features here. 00:55:25.660 --> 00:55:28.930 align:middle line:84% And what we can do is we have our two data sets. 00:55:28.930 --> 00:55:34.020 align:middle line:84% So we'll say the training set addFeatureExtractors 00:55:34.020 --> 00:55:36.660 align:middle line:84% and add all those features extractors 00:55:36.660 --> 00:55:38.570 align:middle line:84% and then same thing for the test set. 00:55:38.570 --> 00:55:43.140 align:middle line:90% 00:55:43.140 --> 00:55:44.797 align:middle line:84% And this, as I said, you don't need 00:55:44.797 --> 00:55:46.380 align:middle line:84% to be typing too much because I'm just 00:55:46.380 --> 00:55:48.250 align:middle line:90% going to go pretty fast. 00:55:48.250 --> 00:55:53.730 align:middle line:84% And what we'll do is I'm just going to cut and paste this 00:55:53.730 --> 00:55:57.420 align:middle line:84% because all we're doing is dividing things up 00:55:57.420 --> 00:56:00.040 align:middle line:84% into either the test or the training set, 00:56:00.040 --> 00:56:02.860 align:middle line:84% depending on if it's even or odd. 00:56:02.860 --> 00:56:06.450 align:middle line:84% And so we'll just add the data on here 00:56:06.450 --> 00:56:08.280 align:middle line:90% with some nice little features. 00:56:08.280 --> 00:56:11.080 align:middle line:84% And this will just generate all the headers for you. 00:56:11.080 --> 00:56:13.500 align:middle line:90% Whoops, what did I do? 00:56:13.500 --> 00:56:16.240 align:middle line:84% Testing set is what I called it before. 00:56:16.240 --> 00:56:18.395 align:middle line:84% Sorry, training and testing, that makes sense. 00:56:18.395 --> 00:56:21.330 align:middle line:90% 00:56:21.330 --> 00:56:22.630 align:middle line:90% Did I do something wrong? 00:56:22.630 --> 00:56:25.300 align:middle line:90% Oh I have to change it up here. 00:56:25.300 --> 00:56:26.370 align:middle line:90% There we go. 00:56:26.370 --> 00:56:29.830 align:middle line:84% Starting all over, hopefully, I don't have any num correct. 00:56:29.830 --> 00:56:33.870 align:middle line:84% Great, and that's just doing all the feature extraction 00:56:33.870 --> 00:56:35.080 align:middle line:90% of getting them in. 00:56:35.080 --> 00:56:37.380 align:middle line:84% That's going to take a little bit of time. 00:56:37.380 --> 00:56:39.570 align:middle line:84% So we're going to switch like they 00:56:39.570 --> 00:56:43.110 align:middle line:84% do in the cooking shows to the already baked version 00:56:43.110 --> 00:56:45.180 align:middle line:90% where it's all done. 00:56:45.180 --> 00:56:47.400 align:middle line:90% OK. 00:56:47.400 --> 00:56:48.990 align:middle line:90% Oops. 00:56:48.990 --> 00:56:51.840 align:middle line:90% Here we go. 00:56:51.840 --> 00:56:53.090 align:middle line:90% Where do we go? 00:56:53.090 --> 00:56:55.660 align:middle line:90% 00:56:55.660 --> 00:56:58.900 align:middle line:84% Yep, so the next couple of things that you do, 00:56:58.900 --> 00:57:02.160 align:middle line:84% you want to process all your training set 00:57:02.160 --> 00:57:04.390 align:middle line:84% and write it out just like before, 00:57:04.390 --> 00:57:05.950 align:middle line:90% process it and write it out. 00:57:05.950 --> 00:57:08.020 align:middle line:90% Why is it taking so much longer? 00:57:08.020 --> 00:57:10.780 align:middle line:84% Because we gave it a whole bunch of features. 00:57:10.780 --> 00:57:13.240 align:middle line:84% And some of them take a long time to do, 00:57:13.240 --> 00:57:17.670 align:middle line:84% like ones that have to do with variability between parts 00:57:17.670 --> 00:57:22.020 align:middle line:84% and parallel fifths, finding if it's a multi-part thing and so 00:57:22.020 --> 00:57:23.460 align:middle line:90% on. 00:57:23.460 --> 00:57:26.430 align:middle line:84% But then the next row ends up being 00:57:26.430 --> 00:57:28.180 align:middle line:90% the exact same type of thing. 00:57:28.180 --> 00:57:29.380 align:middle line:90% We have a majority learner. 00:57:29.380 --> 00:57:31.840 align:middle line:84% We're going to still use the knnLearner 00:57:31.840 --> 00:57:39.110 align:middle line:84% and then try to see how well it did and keep looking. 00:57:39.110 --> 00:57:42.350 align:middle line:84% And again, the majority learner is still 59%. 00:57:42.350 --> 00:57:43.970 align:middle line:90% You hope that doesn't change. 00:57:43.970 --> 00:57:47.390 align:middle line:84% And now, using all of these extra features, 00:57:47.390 --> 00:57:53.710 align:middle line:84% we've gotten our number from 87% all the way up to 61%. 00:57:53.710 --> 00:57:58.510 align:middle line:84% So sometimes, just throwing more feature extractors 00:57:58.510 --> 00:58:01.750 align:middle line:84% at the problem is just going to have you 00:58:01.750 --> 00:58:04.720 align:middle line:90% wait longer at the computer. 00:58:04.720 --> 00:58:09.820 align:middle line:84% Sometimes, thinking really carefully about what you already 00:58:09.820 --> 00:58:13.220 align:middle line:84% think the data might look like is a pretty good thing, 00:58:13.220 --> 00:58:16.030 align:middle line:84% and especially if you're working with a small data 00:58:16.030 --> 00:58:18.430 align:middle line:84% set where it's not going to be able to do 00:58:18.430 --> 00:58:21.490 align:middle line:84% the kinds of deep neural network learning. 00:58:21.490 --> 00:58:23.890 align:middle line:84% You have to keep going through enough 00:58:23.890 --> 00:58:26.930 align:middle line:84% to really build a model of the set. 00:58:26.930 --> 00:58:33.640 align:middle line:84% As some people say that look, the neural networks 00:58:33.640 --> 00:58:35.770 align:middle line:90% are getting better and better. 00:58:35.770 --> 00:58:39.260 align:middle line:84% And we just need to throw more data at it. 00:58:39.260 --> 00:58:41.950 align:middle line:84% But imagine if you wanted to have 00:58:41.950 --> 00:58:46.310 align:middle line:84% a program that was automatically reading handwriting. 00:58:46.310 --> 00:58:51.440 align:middle line:84% And you gave the program all these papers of handwriting. 00:58:51.440 --> 00:58:53.560 align:middle line:84% And you just gave them an audio file 00:58:53.560 --> 00:58:56.840 align:middle line:84% of how it drops, how it sounds when it drops and try to do it. 00:58:56.840 --> 00:58:59.870 align:middle line:84% You're giving it kind of the wrong data to learn on. 00:58:59.870 --> 00:59:03.020 align:middle line:84% And so a lot of the extractors now, 00:59:03.020 --> 00:59:06.010 align:middle line:84% unless you're writing your own or unless you're using really 00:59:06.010 --> 00:59:09.760 align:middle line:84% musically sensitive ones, you can save the computer 00:59:09.760 --> 00:59:14.290 align:middle line:84% a lot of time by trying to already say, hey, let's process 00:59:14.290 --> 00:59:18.880 align:middle line:84% this as a symbols ahead of time or something like that. 00:59:18.880 --> 00:59:23.660 align:middle line:84% A lot of, by the way, state of the art of symbolic music 00:59:23.660 --> 00:59:25.930 align:middle line:90% machine learning things are-- 00:59:25.930 --> 00:59:29.830 align:middle line:84% you all remember the piano roll where 00:59:29.830 --> 00:59:34.220 align:middle line:84% we put things like this to show length and stuff like that? 00:59:34.220 --> 00:59:38.230 align:middle line:84% Yeah, well, it's converting the entire musical score 00:59:38.230 --> 00:59:42.790 align:middle line:84% to a piano roll, saving that as an image, 00:59:42.790 --> 00:59:47.200 align:middle line:84% and then using image classifiers and image generators 00:59:47.200 --> 00:59:50.140 align:middle line:84% to generate pieces that look like the piano rolls 00:59:50.140 --> 00:59:51.770 align:middle line:90% and then translating them back. 00:59:51.770 --> 00:59:56.060 align:middle line:84% So I'm hoping that over the next decade or so, 00:59:56.060 --> 00:59:58.750 align:middle line:84% we can do better and do some things that 00:59:58.750 --> 01:00:02.920 align:middle line:84% help these neural nets and classifiers 01:00:02.920 --> 01:00:05.530 align:middle line:84% work with data that's a little bit closer 01:00:05.530 --> 01:00:08.320 align:middle line:90% to the original numbers. 01:00:08.320 --> 01:00:14.350 align:middle line:84% OK, I have time for just a little fun 01:00:14.350 --> 01:00:21.620 align:middle line:84% thing with my favorite low feature extraction thing. 01:00:21.620 --> 01:00:25.215 align:middle line:84% So let me just show a couple little things. 01:00:25.215 --> 01:00:26.410 align:middle line:90% OK. 01:00:26.410 --> 01:00:28.220 align:middle line:90% Yeah. 01:00:28.220 --> 01:00:31.310 align:middle line:84% First off, the importance of thinking about data sets 01:00:31.310 --> 01:00:34.700 align:middle line:84% and ground truths, this is, I think, 01:00:34.700 --> 01:00:37.430 align:middle line:84% one of the first pieces that showed 01:00:37.430 --> 01:00:40.460 align:middle line:84% deep learning, the kind of for people in the 01:00:40.460 --> 01:00:45.570 align:middle line:84% know moment like ChatGPT has been in the last year happened. 01:00:45.570 --> 01:00:49.580 align:middle line:84% I think it was about 10 years ago when AlphaZero and AlphaGo 01:00:49.580 --> 01:00:54.320 align:middle line:84% suddenly became the world's greatest go and chess players. 01:00:54.320 --> 01:00:56.210 align:middle line:84% And so a lot of people think, well, this 01:00:56.210 --> 01:00:59.360 align:middle line:84% was with generalized knowledge, no rules being taught 01:00:59.360 --> 01:01:02.300 align:middle line:90% except for legal move or not. 01:01:02.300 --> 01:01:07.460 align:middle line:84% But one of the things to think about the difference between 01:01:07.460 --> 01:01:10.670 align:middle line:84% a lot of the things that they were being trained on and what 01:01:10.670 --> 01:01:14.940 align:middle line:84% you have is that AlphaZero, the chess playing computer, played, 01:01:14.940 --> 01:01:18.770 align:middle line:84% what was it, 700,000 games or a million games or something like 01:01:18.770 --> 01:01:19.560 align:middle line:90% that. 01:01:19.560 --> 01:01:23.280 align:middle line:84% And at the end of every single game it played, 01:01:23.280 --> 01:01:26.940 align:middle line:84% it got perfect feedback on how it did-- 01:01:26.940 --> 01:01:30.200 align:middle line:84% you won, you lost, you tied, you won, you lost, you tied. 01:01:30.200 --> 01:01:33.230 align:middle line:84% A lot of the things that we're interested in doing with music 01:01:33.230 --> 01:01:36.500 align:middle line:84% like music recommendation or things 01:01:36.500 --> 01:01:39.180 align:middle line:84% like this, predicting what's the next great hit, 01:01:39.180 --> 01:01:42.780 align:middle line:84% first off, we don't have that correct data. 01:01:42.780 --> 01:01:45.270 align:middle line:84% Maybe what's the Billboard chart or something like that? 01:01:45.270 --> 01:01:48.525 align:middle line:84% But for how good was this piece, because it's going 01:01:48.525 --> 01:01:49.650 align:middle line:90% to be different every time. 01:01:49.650 --> 01:01:51.500 align:middle line:84% It's going to take us three minutes 01:01:51.500 --> 01:01:54.543 align:middle line:84% to listen to if it's computer generating a song. 01:01:54.543 --> 01:01:56.210 align:middle line:84% And then we're going to give it feedback 01:01:56.210 --> 01:01:59.010 align:middle line:84% that we might feel differently about that piece tomorrow. 01:01:59.010 --> 01:02:00.800 align:middle line:84% So we don't give perfect feedback. 01:02:00.800 --> 01:02:02.490 align:middle line:90% And it takes a long time. 01:02:02.490 --> 01:02:05.660 align:middle line:84% So this is one of the reasons I like working 01:02:05.660 --> 01:02:07.920 align:middle line:90% with some of the older things. 01:02:07.920 --> 01:02:08.940 align:middle line:90% OK, we'll go back. 01:02:08.940 --> 01:02:10.790 align:middle line:90% Great. 01:02:10.790 --> 01:02:12.810 align:middle line:90% I don't think I presented this. 01:02:12.810 --> 01:02:14.220 align:middle line:90% No, I presented in other place. 01:02:14.220 --> 01:02:17.000 align:middle line:84% So this is what we'll end for the last 10 minutes 01:02:17.000 --> 01:02:20.060 align:middle line:84% and talking about some of the interesting things that 01:02:20.060 --> 01:02:22.580 align:middle line:90% come out with features. 01:02:22.580 --> 01:02:26.760 align:middle line:84% So how would you distinguish choral music from Bach 01:02:26.760 --> 01:02:28.190 align:middle line:90% from Monteverdi? 01:02:28.190 --> 01:02:32.400 align:middle line:84% So we're doing a lot of Bach in this class. 01:02:32.400 --> 01:02:34.650 align:middle line:84% But we'll get a little bit more into the ear. 01:02:34.650 --> 01:02:36.990 align:middle line:90% Here's Bach. 01:02:36.990 --> 01:02:39.593 align:middle line:90% [MUSIC PLAYING] 01:02:39.593 --> 01:02:40.440 align:middle line:90% 01:02:40.440 --> 01:02:41.180 align:middle line:90% So this is Bach. 01:02:41.180 --> 01:02:51.440 align:middle line:90% 01:02:51.440 --> 01:02:54.890 align:middle line:84% And so then we'll switch to one of the other great composers 01:02:54.890 --> 01:03:00.020 align:middle line:84% about 100 years before Bach, Claudio Monteverdi. 01:03:00.020 --> 01:03:03.949 align:middle line:84% And so, some of you, this might be the first time you hear him. 01:03:03.949 --> 01:03:06.943 align:middle line:90% [NO AUDIO] 01:03:06.943 --> 01:03:27.930 align:middle line:90% 01:03:27.930 --> 01:03:30.515 align:middle line:84% One second with the Back again so we can get back. 01:03:30.515 --> 01:03:35.010 align:middle line:90% 01:03:35.010 --> 01:03:36.390 align:middle line:90% Great. 01:03:36.390 --> 01:03:39.210 align:middle line:84% What's something you heard different 01:03:39.210 --> 01:03:42.512 align:middle line:90% between the two of them? 01:03:42.512 --> 01:03:44.220 align:middle line:84% STUDENT: Monteverdi was a bit more somber 01:03:44.220 --> 01:03:45.413 align:middle line:90% with more of the minor keys. 01:03:45.413 --> 01:03:47.830 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: OK, somber, more minor keys. 01:03:47.830 --> 01:03:49.140 align:middle line:90% We could do a key detection. 01:03:49.140 --> 01:03:49.640 align:middle line:90% Jonathan. 01:03:49.640 --> 01:03:53.040 align:middle line:90% 01:03:53.040 --> 01:03:55.260 align:middle line:84% STUDENT: I mean, Bach had sounded like there 01:03:55.260 --> 01:03:56.710 align:middle line:90% were more people singing. 01:03:56.710 --> 01:03:58.470 align:middle line:90% But it's hard to tell. 01:03:58.470 --> 01:04:00.637 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: Yeah, but it could be. 01:04:00.637 --> 01:04:03.250 align:middle line:84% It could be more people, fuller than good. 01:04:03.250 --> 01:04:05.050 align:middle line:90% Yeah. 01:04:05.050 --> 01:04:07.560 align:middle line:84% STUDENT: Bach is more together where 01:04:07.560 --> 01:04:11.408 align:middle line:84% Monteverdi has a lot of voices going off doing kind of thing. 01:04:11.408 --> 01:04:13.450 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: Bach more together, 01:04:13.450 --> 01:04:15.730 align:middle line:84% homophonic maybe, Monteverdi doing their own thing. 01:04:15.730 --> 01:04:16.230 align:middle line:90% Good. 01:04:16.230 --> 01:04:17.270 align:middle line:90% One last one. 01:04:17.270 --> 01:04:22.050 align:middle line:90% 01:04:22.050 --> 01:04:24.160 align:middle line:84% OK, that's a pretty good place to start. 01:04:24.160 --> 01:04:29.820 align:middle line:84% So I threw everything in, all the kitchen 01:04:29.820 --> 01:04:33.360 align:middle line:84% sink of feature extractors and got 01:04:33.360 --> 01:04:41.850 align:middle line:84% to see what a particular algorithm did that seemed 01:04:41.850 --> 01:04:43.390 align:middle line:90% to work pretty well. 01:04:43.390 --> 01:04:46.390 align:middle line:84% And the first, here's a little quiz. 01:04:46.390 --> 01:04:49.380 align:middle line:84% It's one of the few quizzes that elementary school students 01:04:49.380 --> 01:04:51.995 align:middle line:90% do better than adults at. 01:04:51.995 --> 01:04:53.620 align:middle line:84% So let's see if we can figure this out. 01:04:53.620 --> 01:04:56.808 align:middle line:90% 7802 is 3. 01:04:56.808 --> 01:04:59.050 align:middle line:90% 1065 is 2. 01:04:59.050 --> 01:05:02.010 align:middle line:84% If you've seen it before, then don't answer. 01:05:02.010 --> 01:05:05.910 align:middle line:90% 998 is 4. 01:05:05.910 --> 01:05:14.340 align:middle line:90% And 1773 is 0. 01:05:14.340 --> 01:05:29.050 align:middle line:84% If you're new to this, 1, 2, 3, 3 circles, 1 circle, 2 circles, 01:05:29.050 --> 01:05:31.960 align:middle line:84% 1, 2, 3, 4 circles and no circles. 01:05:31.960 --> 01:05:35.220 align:middle line:84% So I bring that up to say that sometimes the computer is going 01:05:35.220 --> 01:05:40.107 align:middle line:84% to be looking at things that are really, really far out 01:05:40.107 --> 01:05:41.440 align:middle line:90% and trying to figure out things. 01:05:41.440 --> 01:05:44.820 align:middle line:84% So there's all these things that we could be looking at. 01:05:44.820 --> 01:05:49.110 align:middle line:84% I'll bet that the proportion of instruments that are electronic 01:05:49.110 --> 01:05:52.030 align:middle line:84% isn't going to help with Bach versus Monteverdi. 01:05:52.030 --> 01:05:56.190 align:middle line:84% But one of the things I like about a particular classifier 01:05:56.190 --> 01:05:59.860 align:middle line:84% that is not close to the state of the art 01:05:59.860 --> 01:06:01.780 align:middle line:84% or to the getting the best results. 01:06:01.780 --> 01:06:05.220 align:middle line:90% But this tree learner can-- 01:06:05.220 --> 01:06:09.370 align:middle line:84% unlike being a black box, it can actually show what the path-- 01:06:09.370 --> 01:06:11.730 align:middle line:84% so it's of a decision tree, branching tree-- 01:06:11.730 --> 01:06:13.420 align:middle line:90% it took to get there. 01:06:13.420 --> 01:06:17.770 align:middle line:84% So here's how it can say, to begin, 01:06:17.770 --> 01:06:23.380 align:middle line:84% if the note f sharp below middle C is used more than 14% 01:06:23.380 --> 01:06:26.140 align:middle line:84% of the time, the piece is by Bach. 01:06:26.140 --> 01:06:31.700 align:middle line:84% Otherwise, if the top number in the time signature is 3 or less, 01:06:31.700 --> 01:06:33.670 align:middle line:90% it is by Bach. 01:06:33.670 --> 01:06:36.358 align:middle line:90% Otherwise-- I love this one. 01:06:36.358 --> 01:06:38.650 align:middle line:84% If the distance between the highest note in the soprano 01:06:38.650 --> 01:06:40.390 align:middle line:84% and the lowest note of the bass is 01:06:40.390 --> 01:06:44.150 align:middle line:84% less than 2 octaves and a minor sixth, then it's by Bach. 01:06:44.150 --> 01:06:45.740 align:middle line:84% And then there's three more things. 01:06:45.740 --> 01:06:47.920 align:middle line:84% We get to the final one, count the number 01:06:47.920 --> 01:06:49.430 align:middle line:90% of times each note is used. 01:06:49.430 --> 01:06:51.170 align:middle line:84% Find the note that is used the most often 01:06:51.170 --> 01:06:53.170 align:middle line:84% and the note that is used the second most often. 01:06:53.170 --> 01:06:56.260 align:middle line:84% If the second most often used note is used less than 97% 01:06:56.260 --> 01:07:00.430 align:middle line:84% as often as the most commonly used note, the piece is by Bach. 01:07:00.430 --> 01:07:02.210 align:middle line:90% Otherwise, it's by Monteverdi. 01:07:02.210 --> 01:07:07.420 align:middle line:84% And this worked with 99% accuracy. 01:07:07.420 --> 01:07:13.030 align:middle line:84% What I loved about this is that this was the first decision 01:07:13.030 --> 01:07:14.210 align:middle line:90% to distinguish. 01:07:14.210 --> 01:07:19.330 align:middle line:84% And I was just baffled by it until I realized, well, OK, Bach 01:07:19.330 --> 01:07:23.030 align:middle line:84% has more pieces in sharp keys than Monteverdi is. 01:07:23.030 --> 01:07:24.590 align:middle line:90% OK, that would make sense. 01:07:24.590 --> 01:07:29.290 align:middle line:84% Monteverdi was mostly hovering around no sharps or flats or one 01:07:29.290 --> 01:07:30.410 align:middle line:90% flat assigned. 01:07:30.410 --> 01:07:34.060 align:middle line:84% But then it's minor, and there's raised leading tones and stuff 01:07:34.060 --> 01:07:35.480 align:middle line:90% like that, so why not? 01:07:35.480 --> 01:07:40.180 align:middle line:84% Well, the leading tone tends to get raised in the melody, which 01:07:40.180 --> 01:07:42.250 align:middle line:90% tends to be in a higher voice. 01:07:42.250 --> 01:07:46.180 align:middle line:84% The middle area, tenor, will only 01:07:46.180 --> 01:07:48.860 align:middle line:84% appear as f sharp if the key signature is already f sharp. 01:07:48.860 --> 01:07:51.410 align:middle line:84% So it's a way of learning some things like this. 01:07:51.410 --> 01:07:56.530 align:middle line:84% And this particular one ended up being garbage in, garbage out, 01:07:56.530 --> 01:08:00.890 align:middle line:84% that Monteverdi doesn't using modern time signatures. 01:08:00.890 --> 01:08:06.800 align:middle line:84% And so the people who encoded this chose never to use 2-4, 01:08:06.800 --> 01:08:11.510 align:middle line:84% always put it into 4-4, because it also weren't bar lines. 01:08:11.510 --> 01:08:12.830 align:middle line:90% So you could decide how long. 01:08:12.830 --> 01:08:16.540 align:middle line:84% So this ended up being that the computer learned 01:08:16.540 --> 01:08:19.160 align:middle line:84% not the difference between Bach and Monteverdi 01:08:19.160 --> 01:08:22.240 align:middle line:84% but the difference between Bach's editor, modern editor, 01:08:22.240 --> 01:08:24.560 align:middle line:90% and Monteverdi modern editor. 01:08:24.560 --> 01:08:27.279 align:middle line:84% I have lost the reference, but I have 01:08:27.279 --> 01:08:32.170 align:middle line:84% been told about an audio classifier that 01:08:32.170 --> 01:08:34.160 align:middle line:90% tried to classify by genre. 01:08:34.160 --> 01:08:37.840 align:middle line:84% And it was doing very, very well until it turns out 01:08:37.840 --> 01:08:41.319 align:middle line:84% that different genres of music used different microphones 01:08:41.319 --> 01:08:42.439 align:middle line:90% in general. 01:08:42.439 --> 01:08:44.470 align:middle line:84% And it was picking up the difference 01:08:44.470 --> 01:08:45.830 align:middle line:90% between the microphones. 01:08:45.830 --> 01:08:49.460 align:middle line:84% So if you sing country music on a hip hop microphone, 01:08:49.460 --> 01:08:51.910 align:middle line:90% it would classify as hip hop. 01:08:51.910 --> 01:08:54.010 align:middle line:84% So all of these things, we need to be 01:08:54.010 --> 01:08:58.840 align:middle line:84% really careful about what's going on with the potential 01:08:58.840 --> 01:09:00.140 align:middle line:90% for classification. 01:09:00.140 --> 01:09:04.450 align:middle line:84% And so this is where I like this old school-- 01:09:04.450 --> 01:09:07.645 align:middle line:84% imagine this as a tree diagram of things going left and right, 01:09:07.645 --> 01:09:09.020 align:middle line:84% I guess if you turn it sideways-- 01:09:09.020 --> 01:09:13.779 align:middle line:84% these old school feature classifiers a little bit just 01:09:13.779 --> 01:09:18.920 align:middle line:84% to check what might be happening in the latest black box. 01:09:18.920 --> 01:09:24.450 align:middle line:84% So you guys did a remarkable stamina. 01:09:24.450 --> 01:09:29.279 align:middle line:84% You got through the zombie of music eating feature extraction. 01:09:29.279 --> 01:09:34.229 align:middle line:84% And I hope that some people will choose to find-- 01:09:34.229 --> 01:09:37.910 align:middle line:84% if you find this to be helpful in doing something 01:09:37.910 --> 01:09:41.550 align:middle line:84% in your final project, you now have some of the tools to do it. 01:09:41.550 --> 01:09:45.630 align:middle line:84% On Friday, we'll continue with searching and similarity, 01:09:45.630 --> 01:09:49.250 align:middle line:84% and will be giving the first 15 minutes explaining 01:09:49.250 --> 01:09:50.760 align:middle line:90% what he's doing at MIT. 01:09:50.760 --> 01:09:53.029 align:middle line:84% He's a professor of computer science 01:09:53.029 --> 01:09:57.300 align:middle line:84% and what he's been doing with music on his work. 01:09:57.300 --> 01:09:59.343 align:middle line:90% So there'll be a start. 01:09:59.343 --> 01:10:01.260 align:middle line:84% And then I'll continue with some other things. 01:10:01.260 --> 01:10:03.350 align:middle line:90% Great, thanks. 01:10:03.350 --> 01:10:14.000 align:middle line:90%