1
00:00:01,640 --> 00:00:04,040
The following content is
provided under a Creative

2
00:00:04,040 --> 00:00:05,580
Commons license.

3
00:00:05,580 --> 00:00:07,880
Your support will help
MIT OpenCourseWare

4
00:00:07,880 --> 00:00:12,270
continue to offer high quality
educational resources for free.

5
00:00:12,270 --> 00:00:14,870
To make a donation or
view additional materials

6
00:00:14,870 --> 00:00:18,830
from hundreds of MIT courses,
visit MIT OpenCourseWare

7
00:00:18,830 --> 00:00:20,000
at ocw.mit.edu.

8
00:00:21,869 --> 00:00:23,910
MATT WILSON: Now, why are
we interested in sleep?

9
00:00:23,910 --> 00:00:26,330
So, we kind of think about
this as two modes, online mode

10
00:00:26,330 --> 00:00:27,038
and offline mode.

11
00:00:27,038 --> 00:00:29,920
In the online mode, you're
taking information in.

12
00:00:29,920 --> 00:00:32,600
In the offline mode,
you're going back

13
00:00:32,600 --> 00:00:33,737
and evaluating information.

14
00:00:33,737 --> 00:00:35,570
What's the purpose of
evaluating information

15
00:00:35,570 --> 00:00:36,445
that you've taken in?

16
00:00:36,445 --> 00:00:38,528
Well, if we think about
the general problem that's

17
00:00:38,528 --> 00:00:40,400
being solved, the
problem of intelligence.

18
00:00:40,400 --> 00:00:45,602
The problem of intelligence is
trying to understand and infer

19
00:00:45,602 --> 00:00:47,810
these sort of deep generalizable
relationships rules.

20
00:00:47,810 --> 00:00:50,210
You're trying to extract
rules from instances.

21
00:00:50,210 --> 00:00:52,190
And you'd like those
rules to be as generally

22
00:00:52,190 --> 00:00:54,840
applicable as possible.

23
00:00:54,840 --> 00:00:57,800
And in order to do
that, presumably, one

24
00:00:57,800 --> 00:01:02,030
has to go back and evaluate
the many individual instances

25
00:01:02,030 --> 00:01:05,000
to try to extract some kind
of statistical regularity,

26
00:01:05,000 --> 00:01:08,210
and then perhaps evaluate
models that you have constructed

27
00:01:08,210 --> 00:01:11,250
in terms of their consistency
with individual instances

28
00:01:11,250 --> 00:01:12,500
that you've already collected.

29
00:01:12,500 --> 00:01:15,470
Or perhaps future instances
that you haven't yet.

30
00:01:15,470 --> 00:01:20,974
So you have the raw material,
you build a little model,

31
00:01:20,974 --> 00:01:22,640
and then you continually
test that model

32
00:01:22,640 --> 00:01:25,040
against new information
that comes in.

33
00:01:25,040 --> 00:01:26,790
And the question is,
when can you do that?

34
00:01:26,790 --> 00:01:29,840
Now, you could do
that when you're

35
00:01:29,840 --> 00:01:32,740
out and about in the world.

36
00:01:32,740 --> 00:01:35,780
As you saw and read in that
paper, when animals are sitting

37
00:01:35,780 --> 00:01:37,580
quietly, they very
quickly can switch

38
00:01:37,580 --> 00:01:40,160
into this kind of offline mode
that looks a lot like sleep.

39
00:01:40,160 --> 00:01:43,235
In fact, electrophysiologically,
sleep and quiet wakefulness

40
00:01:43,235 --> 00:01:46,280
in the hippocampus are
nearly indistinguishable.

41
00:01:46,280 --> 00:01:47,870
It's the same kind
of offline mode.

42
00:01:47,870 --> 00:01:51,920
It says, OK, when the
hippocampus is not

43
00:01:51,920 --> 00:01:54,380
being used to take
new information in,

44
00:01:54,380 --> 00:01:58,120
I quickly switch into this
offline evaluation mode.

45
00:01:58,120 --> 00:02:03,200
But during sleep, I no
longer have the constraint

46
00:02:03,200 --> 00:02:05,500
of having to direct behavior.

47
00:02:05,500 --> 00:02:08,220
Behavior's shut off,
inputs are shut off,

48
00:02:08,220 --> 00:02:11,420
and now I can switch into this
purely internal introspective

49
00:02:11,420 --> 00:02:11,930
mode.

50
00:02:11,930 --> 00:02:14,280
So what goes on during sleep.

51
00:02:14,280 --> 00:02:17,510
In this simple experiment,
we look at the activity

52
00:02:17,510 --> 00:02:19,950
when the animal is
performing behavioral tasks.

53
00:02:19,950 --> 00:02:22,070
And then we examine
activity during sleep,

54
00:02:22,070 --> 00:02:23,810
both before and after,
and ask, is there

55
00:02:23,810 --> 00:02:27,020
anything about
behavioral experience

56
00:02:27,020 --> 00:02:30,530
that changes activity
during sleep?

57
00:02:30,530 --> 00:02:34,514
And what we found was that
if you look at activity

58
00:02:34,514 --> 00:02:35,930
during behavior
in which you think

59
00:02:35,930 --> 00:02:39,120
about these spatial
sequences being expressed,

60
00:02:39,120 --> 00:02:40,620
and you look during
sleep afterward,

61
00:02:40,620 --> 00:02:42,953
you find that the spatial
sequences are expressed again.

62
00:02:42,953 --> 00:02:48,320
So the hippocampus replays the
firing of these cell sequences.

63
00:02:48,320 --> 00:02:50,970
But it replays the firing
of these cell sequences

64
00:02:50,970 --> 00:02:54,284
at a time scale that appears
to be compressed relative

65
00:02:54,284 --> 00:02:55,450
to the behavioral timescale.

66
00:02:55,450 --> 00:02:58,130
So these are eight place cells.

67
00:02:58,130 --> 00:03:00,200
Animals were walking
from left to right.

68
00:03:00,200 --> 00:03:02,790
The ticks indicate that's
the location of the peak.

69
00:03:02,790 --> 00:03:06,290
So these place cells will
fire one through eight

70
00:03:06,290 --> 00:03:08,791
as the animal moves along the
track over about five seconds.

71
00:03:08,791 --> 00:03:10,248
That's how long it
takes the animal

72
00:03:10,248 --> 00:03:11,450
to walk from left to right.

73
00:03:11,450 --> 00:03:14,180
The same sequence of
1, 2, 3, 4, 5, 6, 7, 8

74
00:03:14,180 --> 00:03:16,130
gets replayed during
sleep, but now

75
00:03:16,130 --> 00:03:19,730
over about 150 milliseconds.

76
00:03:19,730 --> 00:03:23,390
Same sequence, so you're
preserving time order,

77
00:03:23,390 --> 00:03:24,330
not absolute time.

78
00:03:27,080 --> 00:03:32,150
And that when you ask, when
these sequences are expressed,

79
00:03:32,150 --> 00:03:35,150
what's going on in the
local field potential

80
00:03:35,150 --> 00:03:36,450
in these oscillations?

81
00:03:36,450 --> 00:03:40,910
And this is where you find
these sharp wave ripple events.

82
00:03:40,910 --> 00:03:43,730
This is this ripple-like event
that I was describing to you.

83
00:03:43,730 --> 00:03:46,820
So these sharp wave
ripples are when

84
00:03:46,820 --> 00:03:49,850
these apparently compressed
sequences are being expressed.

85
00:03:49,850 --> 00:03:53,670
I could point out that a simple
model for these sharp wave

86
00:03:53,670 --> 00:03:58,850
ripple reactivated sequences
is the same model that I like

87
00:03:58,850 --> 00:04:01,280
to use to explain
phase precession

88
00:04:01,280 --> 00:04:02,480
during the fade oscillation.

89
00:04:02,480 --> 00:04:06,380
That is, I give it an
input, I sweep inhibition

90
00:04:06,380 --> 00:04:09,890
from high to low such that cells
that are getting the strongest

91
00:04:09,890 --> 00:04:11,840
input fire earlier.

92
00:04:11,840 --> 00:04:14,990
And this will
reactivate a sequence.

93
00:04:14,990 --> 00:04:17,870
So the difference
between a theta sequence

94
00:04:17,870 --> 00:04:20,220
and a reactivated
sequence is really just,

95
00:04:20,220 --> 00:04:21,529
where's the input coming from.

96
00:04:21,529 --> 00:04:25,160
If the input's coming from what
I'm actually experiencing right

97
00:04:25,160 --> 00:04:27,800
now, now it's a theta sequence.

98
00:04:27,800 --> 00:04:30,590
And as I move, the input
changes systematically.

99
00:04:30,590 --> 00:04:34,400
And so, again, it looks like
it's encoding information

100
00:04:34,400 --> 00:04:35,810
about space.

101
00:04:35,810 --> 00:04:38,990
If I'm now offline,
and this information

102
00:04:38,990 --> 00:04:41,270
is being delivered
to the hippocampus

103
00:04:41,270 --> 00:04:43,970
from some other
source, you can apply

104
00:04:43,970 --> 00:04:47,570
exactly the same operation,
get the same kind of sequence.

105
00:04:47,570 --> 00:04:50,070
It's just now it's
on information

106
00:04:50,070 --> 00:04:57,855
that is not tied to your
immediate context or location.

107
00:04:57,855 --> 00:04:58,980
That's the only difference.

108
00:04:58,980 --> 00:05:00,271
Where is the input coming from?

109
00:05:02,510 --> 00:05:07,160
Now further, if you
think about this model,

110
00:05:07,160 --> 00:05:10,579
you can imagine the
depth of disinhibition

111
00:05:10,579 --> 00:05:12,620
could actually affect the
length of the sequence.

112
00:05:12,620 --> 00:05:15,120
But that's just sort
of a mechanistic thing.

113
00:05:15,120 --> 00:05:18,050
And if you actually look
at the mechanisms that

114
00:05:18,050 --> 00:05:21,090
regulate sharp wave ripples
in the theta oscillation,

115
00:05:21,090 --> 00:05:23,880
it basically comes from
the same structure.

116
00:05:23,880 --> 00:05:27,530
It's a structure that
regulates acetylcholine,

117
00:05:27,530 --> 00:05:30,380
and it's a neuromodulator
that's associated

118
00:05:30,380 --> 00:05:32,750
with attention and
memory, and the structure

119
00:05:32,750 --> 00:05:33,800
called the medial septum.

120
00:05:33,800 --> 00:05:36,560
So the medial septum
provides the drive,

121
00:05:36,560 --> 00:05:38,330
the cholinergic drive,
of the hippocampus.

122
00:05:38,330 --> 00:05:42,260
Damage to the medial septum,
loss of cholinergic tone,

123
00:05:42,260 --> 00:05:48,500
was one of the dominant models
of neurodegenerative cognitive

124
00:05:48,500 --> 00:05:50,530
and memory loss in
Alzheimer's disease.

125
00:05:50,530 --> 00:05:53,060
So what you find is
one of the earliest

126
00:05:53,060 --> 00:05:57,890
indications of neurodegenerative
damage in Alzheimer's is

127
00:05:57,890 --> 00:05:59,450
the loss of cholinergic tone.

128
00:05:59,450 --> 00:06:03,350
And the systems that
begin to break down

129
00:06:03,350 --> 00:06:06,110
are the systems that
actually fall along

130
00:06:06,110 --> 00:06:10,260
this limbic pathway starting
with hippocampus-entorhinal

131
00:06:10,260 --> 00:06:10,760
cortex.

132
00:06:10,760 --> 00:06:15,320
So it's as though the
cholinergic system that

133
00:06:15,320 --> 00:06:22,290
regulates the expression of this
oscillation in the hippocampus,

134
00:06:22,290 --> 00:06:24,410
when it breaks down, it
leads to general memory

135
00:06:24,410 --> 00:06:26,030
loss and disruption.

136
00:06:26,030 --> 00:06:28,460
And it turns out the
medial septum is also

137
00:06:28,460 --> 00:06:31,460
involved in regulating the
expression of these sharp wave

138
00:06:31,460 --> 00:06:32,100
ripples.

139
00:06:32,100 --> 00:06:35,950
So, same system,
different modes.

140
00:06:35,950 --> 00:06:40,460
One quite active, online.

141
00:06:40,460 --> 00:06:42,290
One inactive, offline.

142
00:06:42,290 --> 00:06:45,680
Same kind of modulation
of excitability

143
00:06:45,680 --> 00:06:46,940
through inhibition.

144
00:06:46,940 --> 00:06:48,620
But that's the idea.

145
00:06:48,620 --> 00:06:52,040
Simple model, sweep inhibition,
that gives you the sequences.

146
00:06:52,040 --> 00:06:53,860
And then if you can
control the input,

147
00:06:53,860 --> 00:06:56,150
you control the input
to control the content,

148
00:06:56,150 --> 00:06:57,710
you control the
inhibition to control

149
00:06:57,710 --> 00:06:59,030
the structure of the timing.

150
00:06:59,030 --> 00:07:00,330
Those are the two things.

151
00:07:00,330 --> 00:07:02,496
So the question is, how
could you control the input?

152
00:07:05,284 --> 00:07:06,950
Well, you're kind of
thinking about what

153
00:07:06,950 --> 00:07:08,886
is the input into
the hippocampus

154
00:07:08,886 --> 00:07:10,010
under these two conditions.

155
00:07:10,010 --> 00:07:12,147
And as I mentioned,
hippocampus, you've

156
00:07:12,147 --> 00:07:13,230
got the entorhinal cortex.

157
00:07:13,230 --> 00:07:15,396
Entorhinal cortex gets
information across the brain.

158
00:07:15,396 --> 00:07:17,840
It's sort of these
sensory association areas,

159
00:07:17,840 --> 00:07:19,460
visual cortex, auditory cortex.

160
00:07:19,460 --> 00:07:21,530
So all this information
about the world

161
00:07:21,530 --> 00:07:26,120
converging on the hippocampus
and then getting modulated.

162
00:07:26,120 --> 00:07:29,110
And so let's look in, for
instance, the visual cortex.

163
00:07:29,110 --> 00:07:30,860
So if we simultaneous
record visual cortex

164
00:07:30,860 --> 00:07:33,500
and the hippocampus, we could
see how these two structures

165
00:07:33,500 --> 00:07:34,940
communicate.

166
00:07:34,940 --> 00:07:37,460
And when we do this
in a simple task--

167
00:07:37,460 --> 00:07:39,450
this is like a little
figure eight task--

168
00:07:39,450 --> 00:07:41,270
and I won't go into
a lot of the details.

169
00:07:41,270 --> 00:07:42,895
But one interesting
thing that came out

170
00:07:42,895 --> 00:07:44,436
from doing this
experiment, recording

171
00:07:44,436 --> 00:07:46,070
the visual cortex
and the hippocampus,

172
00:07:46,070 --> 00:07:49,130
is that recording in the visual
cortex when an animal is moving

173
00:07:49,130 --> 00:07:53,810
in space, what you find is
cells in the visual cortex

174
00:07:53,810 --> 00:07:55,850
have spatial-like
receptive fields.

175
00:07:55,850 --> 00:07:58,460
Similar to the hippocampus,
though not as spatially tuned.

176
00:07:58,460 --> 00:08:03,980
But here, for instance, are
eight visual cortical cells.

177
00:08:03,980 --> 00:08:08,381
And you can see that
they fire in a sequence.

178
00:08:08,381 --> 00:08:10,380
They'll have different
spatial receptive fields.

179
00:08:10,380 --> 00:08:14,121
This shows where this one visual
cortical cell likes to fire.

180
00:08:14,121 --> 00:08:15,620
Different visual
cortical cells will

181
00:08:15,620 --> 00:08:17,460
fire at different locations.

182
00:08:17,460 --> 00:08:19,250
So when animals are
moving in space,

183
00:08:19,250 --> 00:08:21,560
you actually see
sequential activation

184
00:08:21,560 --> 00:08:24,230
of these visual
cortical responses.

185
00:08:24,230 --> 00:08:29,590
And so if you have visual
perceptual sequences

186
00:08:29,590 --> 00:08:31,900
in hippocampal
spatial sequences,

187
00:08:31,900 --> 00:08:35,830
one question is, how
do those sequences

188
00:08:35,830 --> 00:08:39,409
relate offline when the animal,
for instance, during sleep.

189
00:08:39,409 --> 00:08:41,220
So you have sequences
during sleep

190
00:08:41,220 --> 00:08:42,970
in the hippocampus,
sequences during sleep

191
00:08:42,970 --> 00:08:44,650
in the visual cortex.

192
00:08:44,650 --> 00:08:49,250
And it turns out that those two
things are actually correlated.

193
00:08:49,250 --> 00:08:54,300
So when hippocampus plays
out a spatial sequence,

194
00:08:54,300 --> 00:08:56,950
the visual cortex plays
out a visual sequence

195
00:08:56,950 --> 00:09:02,500
that corresponds to the visual
responses at those locations.

196
00:09:02,500 --> 00:09:04,880
So hippocampus
plays out a sequence

197
00:09:04,880 --> 00:09:06,970
of where it was,
visual cortex expresses

198
00:09:06,970 --> 00:09:11,350
the visual stimuli that were
present along that sequence.

199
00:09:11,350 --> 00:09:13,660
One thing about
these sequences when

200
00:09:13,660 --> 00:09:17,230
we're looking at hippocampal
neocortical interactions

201
00:09:17,230 --> 00:09:20,590
is that the sequences are now
at a much longer time scale.

202
00:09:20,590 --> 00:09:22,760
And in fact this
time scale, which

203
00:09:22,760 --> 00:09:26,950
here is on the order of about
half a second to a second,

204
00:09:26,950 --> 00:09:29,260
corresponds to
another oscillation

205
00:09:29,260 --> 00:09:32,540
that is characteristic of sleep
known as the slow oscillation.

206
00:09:32,540 --> 00:09:35,620
So when you go to
sleep, brain rhythm,

207
00:09:35,620 --> 00:09:39,250
the oscillations, the dominant
frequencies start to slow down.

208
00:09:39,250 --> 00:09:42,700
And you'll get this
oscillation in this one

209
00:09:42,700 --> 00:09:46,960
hertz or so frequency range.

210
00:09:46,960 --> 00:09:49,030
If you record activity
from a bunch of cells,

211
00:09:49,030 --> 00:09:49,840
it looks like this.

212
00:09:49,840 --> 00:09:51,400
This is the activity
of a whole bunch

213
00:09:51,400 --> 00:09:54,010
of cells in the visual
cortex in the hippocampus.

214
00:09:54,010 --> 00:09:58,250
And you see that the activity is
flipping between lots of cells

215
00:09:58,250 --> 00:10:00,380
active, no cells active.

216
00:10:00,380 --> 00:10:01,900
These are the
so-called up and down

217
00:10:01,900 --> 00:10:06,220
states of cells during sleep.

218
00:10:06,220 --> 00:10:09,850
That every half a second to
a second or so many cells

219
00:10:09,850 --> 00:10:11,950
will become active, then
they'll all be shut off,

220
00:10:11,950 --> 00:10:13,324
then they'll become
active again.

221
00:10:13,324 --> 00:10:16,759
So you are flipping between
these up and down-like states.

222
00:10:16,759 --> 00:10:18,550
And if you look at
these up and down states

223
00:10:18,550 --> 00:10:20,590
in the visual
cortex, and you look

224
00:10:20,590 --> 00:10:22,390
at similar activity
in the hippocampus,

225
00:10:22,390 --> 00:10:24,730
you see these up and down-like
states in the hippocampus

226
00:10:24,730 --> 00:10:25,229
as well.

227
00:10:25,229 --> 00:10:28,480
But you'll notice that the
up state in the neocortex

228
00:10:28,480 --> 00:10:30,890
leads up state in
the hippocampus.

229
00:10:30,890 --> 00:10:33,070
So neocortex first, then
hippocampus, neocortex,

230
00:10:33,070 --> 00:10:34,630
hippocampus.

231
00:10:34,630 --> 00:10:36,490
Neocortex seems to lead.

232
00:10:36,490 --> 00:10:40,000
So again, this question of,
who's providing the input?

233
00:10:40,000 --> 00:10:43,030
That simple model where
I'm sweeping inhibition

234
00:10:43,030 --> 00:10:46,870
during the theta, during active
behavior, that information

235
00:10:46,870 --> 00:10:49,247
is coming in from perception,
is what I'm saying.

236
00:10:49,247 --> 00:10:51,580
During sleep, the question
is, where is that information

237
00:10:51,580 --> 00:10:52,670
coming from?

238
00:10:52,670 --> 00:10:55,310
Well, this would say, this
is where it's coming from.

239
00:10:55,310 --> 00:10:56,920
It's coming from
these cortical areas.

240
00:10:56,920 --> 00:11:00,975
Let's say sensory, visual
cortical areas turn on.

241
00:11:00,975 --> 00:11:02,975
They provide the information
to the hippocampus.

242
00:11:02,975 --> 00:11:05,500
Now the hippocampus turns
on, and the hippocampus

243
00:11:05,500 --> 00:11:09,175
responding to input coming
in from the sensory cortex.

244
00:11:09,175 --> 00:11:11,950
Now, if that's
the case, could we

245
00:11:11,950 --> 00:11:14,260
manipulate hippocampal
activity by manipulating

246
00:11:14,260 --> 00:11:15,550
the sensory cortex?

247
00:11:15,550 --> 00:11:19,550
And so this experiment, simple
experiment, answer was yes.

248
00:11:19,550 --> 00:11:21,716
In this case we use
the auditory system.

249
00:11:21,716 --> 00:11:23,590
One of the reasons to
use the auditory system

250
00:11:23,590 --> 00:11:25,006
is that, unlike
the visual system,

251
00:11:25,006 --> 00:11:28,400
the auditory system remains in
a state of persistent vigilance,

252
00:11:28,400 --> 00:11:29,257
even during sleep.

253
00:11:29,257 --> 00:11:30,840
Auditory cortical
responses, even when

254
00:11:30,840 --> 00:11:33,540
the animal goes
to sleep, measure

255
00:11:33,540 --> 00:11:35,740
auditory cortical responses
same as when it's awake.

256
00:11:35,740 --> 00:11:38,680
So essentially the
auditory cortex

257
00:11:38,680 --> 00:11:42,490
stays vigilant
during sleep, which

258
00:11:42,490 --> 00:11:44,510
has clear evolutionary value.

259
00:11:44,510 --> 00:11:47,890
Animal's asleep, it's
trying to minimize arousal

260
00:11:47,890 --> 00:11:52,390
by shutting off visual input,
which may be of limited value

261
00:11:52,390 --> 00:11:56,140
anyway, given the
circadian nocturnal nature

262
00:11:56,140 --> 00:11:59,012
of visual stimuli.

263
00:11:59,012 --> 00:12:00,970
So if you can't see
anything, you might as well

264
00:12:00,970 --> 00:12:02,011
actually close your eyes.

265
00:12:02,011 --> 00:12:03,460
But you can still hear things.

266
00:12:03,460 --> 00:12:06,640
So you and other
animals are still

267
00:12:06,640 --> 00:12:09,786
listening to pick up
threats that might require

268
00:12:09,786 --> 00:12:10,910
that they actually wake up.

269
00:12:10,910 --> 00:12:12,880
So anyway, taking
advantage of that.

270
00:12:12,880 --> 00:12:16,690
Auditory system on, so
train an animal on task

271
00:12:16,690 --> 00:12:21,520
where it learns to associate
auditory cues with locations.

272
00:12:21,520 --> 00:12:23,700
Right sound means go
over here to get food,

273
00:12:23,700 --> 00:12:25,540
left sound means go
over there to get food.

274
00:12:25,540 --> 00:12:26,856
Very simple task.

275
00:12:26,856 --> 00:12:28,480
Then animal goes to
sleep, and you just

276
00:12:28,480 --> 00:12:29,710
continue to play the sounds.

277
00:12:29,710 --> 00:12:31,360
So it's learned something.

278
00:12:31,360 --> 00:12:36,040
And now you try to bias cortical
activity during sleep and ask,

279
00:12:36,040 --> 00:12:38,161
what does that do?

280
00:12:38,161 --> 00:12:39,160
And so this is the idea.

281
00:12:39,160 --> 00:12:40,659
When the animal is
actually running,

282
00:12:40,659 --> 00:12:42,179
we can decode
hippocampal activity.

283
00:12:42,179 --> 00:12:44,470
So we can tell, oh, here's
the hippocampal pattern that

284
00:12:44,470 --> 00:12:48,920
corresponds to the left
side or the right side.

285
00:12:48,920 --> 00:12:51,850
And in this case, when the
animal's performing the task,

286
00:12:51,850 --> 00:12:53,410
play the right-hand
sound, animal

287
00:12:53,410 --> 00:12:55,014
goes to the right-hand
side, see this

288
00:12:55,014 --> 00:12:56,180
in the hippocampal response.

289
00:12:56,180 --> 00:13:00,000
Play the left-hand sound, animal
goes to the left-hand side.

290
00:13:00,000 --> 00:13:01,702
Again, the left-hand
place fields

291
00:13:01,702 --> 00:13:03,660
when the animal's on the
left, right-hand place

292
00:13:03,660 --> 00:13:04,493
fields on the right.

293
00:13:04,493 --> 00:13:07,120
So this is what the
experiment and the behavior

294
00:13:07,120 --> 00:13:09,730
looks like from the
standpoint of the hippocampus.

295
00:13:09,730 --> 00:13:11,830
Very clear, right?

296
00:13:11,830 --> 00:13:14,410
Right sound, right
hippocampus, right place cells.

297
00:13:14,410 --> 00:13:16,730
Left sound, left place cells.

298
00:13:16,730 --> 00:13:18,480
But now the animal's
going to go to sleep,

299
00:13:18,480 --> 00:13:19,840
and we're going to
do the same thing.

300
00:13:19,840 --> 00:13:21,880
Continue to play the right
sounds and the left sounds.

301
00:13:21,880 --> 00:13:23,200
It's just that now the
animal is not actually

302
00:13:23,200 --> 00:13:24,070
moving on the track.

303
00:13:24,070 --> 00:13:26,540
And so we ask,
does that-- can we

304
00:13:26,540 --> 00:13:32,260
bias the reactivated
sequences that you get?

305
00:13:32,260 --> 00:13:33,967
So there's the same thing now.

306
00:13:33,967 --> 00:13:36,050
Animal's awake, but now
it's going to go to sleep,

307
00:13:36,050 --> 00:13:37,370
and we just keep
playing the sounds.

308
00:13:37,370 --> 00:13:39,578
The little tics there indicate
the delivery of sounds

309
00:13:39,578 --> 00:13:41,540
every 10 seconds or so.

310
00:13:41,540 --> 00:13:44,690
So we play a sound, and now
we look at the response.

311
00:13:44,690 --> 00:13:47,570
The difference here is that
the animal's not running,

312
00:13:47,570 --> 00:13:48,330
it's not behaving.

313
00:13:48,330 --> 00:13:50,180
Sound, response.

314
00:13:50,180 --> 00:13:53,160
Now we'll decode the activity.

315
00:13:53,160 --> 00:13:56,330
So here, for instance, we take
a little short window here,

316
00:13:56,330 --> 00:13:57,620
about half a second.

317
00:13:57,620 --> 00:14:00,710
And what you see is this
is this up-like state.

318
00:14:00,710 --> 00:14:03,050
Multi-unit activity,
lots of cells firing.

319
00:14:03,050 --> 00:14:04,340
Decode activity.

320
00:14:04,340 --> 00:14:06,980
Now ask when you
play the left sound,

321
00:14:06,980 --> 00:14:08,540
what does activity decode to?

322
00:14:08,540 --> 00:14:13,350
And here you see it's going to
decode into the left-hand side.

323
00:14:13,350 --> 00:14:17,510
So that's the basic hypothesis.

324
00:14:17,510 --> 00:14:19,670
Play the left sound,
you get the replay

325
00:14:19,670 --> 00:14:21,730
of the left side of the track.

326
00:14:21,730 --> 00:14:24,770
You play the right sound,
you get reactivation

327
00:14:24,770 --> 00:14:26,360
of the right side of the track.

328
00:14:26,360 --> 00:14:29,800
And that's what you get.

329
00:14:29,800 --> 00:14:32,170
So when you play
the left-hand sound,

330
00:14:32,170 --> 00:14:35,230
left sound bias activates
place cells on the left side.

331
00:14:35,230 --> 00:14:40,044
Right side bias biases
place cells on the right.

332
00:14:40,044 --> 00:14:42,460
And you can do the same kind
of psychophysical experiments

333
00:14:42,460 --> 00:14:46,480
which has been done in humans,
either with different sensor

334
00:14:46,480 --> 00:14:48,550
modalities-- it's been
done in olfaction.

335
00:14:48,550 --> 00:14:50,380
It's also been done in audition.

336
00:14:50,380 --> 00:14:52,930
So the equivalent
experiment, using auditory

337
00:14:52,930 --> 00:14:56,510
cueing in humans, where you have
people learn the simple task.

338
00:14:56,510 --> 00:14:58,310
And this was done
in Ken Paller's lab

339
00:14:58,310 --> 00:15:00,780
where they do the simple
spatial matching game.

340
00:15:00,780 --> 00:15:02,860
It's like, you know,
where's the cat card,

341
00:15:02,860 --> 00:15:04,037
where's the teapot card?

342
00:15:04,037 --> 00:15:06,370
The variant here is when they
flipped over the cat card,

343
00:15:06,370 --> 00:15:07,985
they would play the
cat sound, a meow.

344
00:15:07,985 --> 00:15:09,526
When they flip over
the teapot sound,

345
00:15:09,526 --> 00:15:16,240
they play a little associated
auditory cue, teapot whistle.

346
00:15:16,240 --> 00:15:18,700
And then people go to sleep.

347
00:15:18,700 --> 00:15:21,490
And during sleep, they would
play either the cat sound

348
00:15:21,490 --> 00:15:22,549
or the teapot sound.

349
00:15:22,549 --> 00:15:24,340
And they found that
when they would wake up

350
00:15:24,340 --> 00:15:27,460
and now they do the task,
if they played the cat

351
00:15:27,460 --> 00:15:29,320
sound during sleep,
they were better

352
00:15:29,320 --> 00:15:33,320
at remembering the location
of the cats than the whistle.

353
00:15:33,320 --> 00:15:36,970
So this says not only does sleep
actually contribute to memory,

354
00:15:36,970 --> 00:15:38,120
but it's selective.

355
00:15:38,120 --> 00:15:41,080
And not only is it selective,
but it can be influenced.

356
00:15:41,080 --> 00:15:42,010
It can be biased.

357
00:15:42,010 --> 00:15:45,967
You can direct the
nature memory processing.

358
00:15:45,967 --> 00:15:47,800
And then our experiments
suggest that, well,

359
00:15:47,800 --> 00:15:52,060
one of the consequences of
this kind of sleep manipulation

360
00:15:52,060 --> 00:15:55,630
would be to bias the
memory reactivation

361
00:15:55,630 --> 00:15:57,580
in the hippocampus.

362
00:15:57,580 --> 00:15:59,520
And so the idea is simple.

363
00:15:59,520 --> 00:16:06,037
That is that cortex
biases the state

364
00:16:06,037 --> 00:16:07,120
that the hippocampus gets.

365
00:16:07,120 --> 00:16:12,250
And then the hippocampus takes
that state, sweeps inhibition,

366
00:16:12,250 --> 00:16:14,820
replays a sequence,
and that sequence then

367
00:16:14,820 --> 00:16:16,550
gets played back to the cortex.

368
00:16:16,550 --> 00:16:21,190
So the cortex, it
sort of knows--

369
00:16:21,190 --> 00:16:24,130
it has these sort
of discrete states

370
00:16:24,130 --> 00:16:27,310
that doesn't necessarily know
what the causal correlations

371
00:16:27,310 --> 00:16:28,060
might be.

372
00:16:28,060 --> 00:16:29,354
What happens next?

373
00:16:29,354 --> 00:16:31,270
It doesn't necessarily
know what happens next.

374
00:16:31,270 --> 00:16:32,228
It knows what happened.

375
00:16:32,228 --> 00:16:34,860
Hippocampus knows
what happened next.

376
00:16:34,860 --> 00:16:36,835
It has lots of instances
of that though.

377
00:16:36,835 --> 00:16:42,320
Well, I saw a red light,
what happens next?

378
00:16:42,320 --> 00:16:43,970
You say, oh, you
know, I saw red light,

379
00:16:43,970 --> 00:16:45,700
and all the cars stopped.

380
00:16:45,700 --> 00:16:46,600
OK, that's great.

381
00:16:46,600 --> 00:16:49,180
If I'm just the cortex,
that's what I learned.

382
00:16:49,180 --> 00:16:50,340
All the car stop.

383
00:16:50,340 --> 00:16:52,798
If I'm the hippocampus, I'm a
little bit smarter than that.

384
00:16:52,798 --> 00:16:57,580
I say, you know, last Tuesday I
was there, there's a red light,

385
00:16:57,580 --> 00:17:00,460
and all the cars, except
for these cyclists.

386
00:17:00,460 --> 00:17:02,780
Man, they didn't stop,
they just kept on going.

387
00:17:02,780 --> 00:17:05,859
So wait a minute,
there's a rule.

388
00:17:05,859 --> 00:17:09,670
Red light, cars stop,
bicycles don't stop.

389
00:17:09,670 --> 00:17:15,640
So you have to refine what
appear to be simple rules.

390
00:17:15,640 --> 00:17:18,190
Often that's actually
used in an example.

391
00:17:18,190 --> 00:17:20,109
How do you use the
prefrontal cortex?

392
00:17:20,109 --> 00:17:23,500
Oh, red light means stop,
green light means go.

393
00:17:23,500 --> 00:17:27,619
That's great, except in the real
world, that rule is too simple.

394
00:17:27,619 --> 00:17:31,340
It has to be refined based upon
your particular experience.

395
00:17:31,340 --> 00:17:33,600
In fact, if you're
really sophisticated,

396
00:17:33,600 --> 00:17:36,940
red light means cars, except
for cabbies, will stop, right?

397
00:17:36,940 --> 00:17:39,600
If I see a cabbie, guaranteed
that guy's not going to stop,

398
00:17:39,600 --> 00:17:42,310
he's going to accelerate, right?

399
00:17:42,310 --> 00:17:44,809
And that's the kind of
information-- that's

400
00:17:44,809 --> 00:17:45,850
what the hippocampus has.

401
00:17:45,850 --> 00:17:47,230
And so you imagine
that's what's going on.

402
00:17:47,230 --> 00:17:49,340
Neocortex, in each one of
these slow oscillations,

403
00:17:49,340 --> 00:17:50,720
is saying red light.

404
00:17:50,720 --> 00:17:54,340
And the hippocampus says,
oh, yeah, OK, cars stop.

405
00:17:54,340 --> 00:17:55,120
Red light again.

406
00:17:55,120 --> 00:17:57,910
Well, there was that bicycle
thing, bicycles continue to go.

407
00:17:57,910 --> 00:17:58,420
Red light.

408
00:17:58,420 --> 00:17:59,950
Well, that was the
cabbie incident.

409
00:17:59,950 --> 00:18:04,720
So now you have all these
sort of causal sequences

410
00:18:04,720 --> 00:18:06,970
that are being expressed
back to the neocortex,

411
00:18:06,970 --> 00:18:12,370
presumably in order to establish
this more comprehensive,

412
00:18:12,370 --> 00:18:17,740
consolidated model of
real world traffic lights,

413
00:18:17,740 --> 00:18:19,870
rather than the
cartoon traffic lights.

414
00:18:19,870 --> 00:18:25,460
And so that would be the idea
of what's going on during sleep.

415
00:18:25,460 --> 00:18:27,200
Now, during quiet
wakefulness you

416
00:18:27,200 --> 00:18:29,490
can think of the
same sort of idea.

417
00:18:29,490 --> 00:18:32,600
And that is that you're
kind of evaluating

418
00:18:32,600 --> 00:18:34,940
multiple casual contingencies.

419
00:18:34,940 --> 00:18:38,030
Each of which might be
expressible as a simple rule,

420
00:18:38,030 --> 00:18:47,240
but might also be experienced
as distinct variations

421
00:18:47,240 --> 00:18:47,930
of that rule.

422
00:18:47,930 --> 00:18:51,110
And the idea is, OK,
do we use the rule,

423
00:18:51,110 --> 00:18:55,310
or do we use the exceptions,
or how do we actually

424
00:18:55,310 --> 00:18:58,507
refine the rule based upon
these different instances

425
00:18:58,507 --> 00:18:59,090
or exceptions?

426
00:18:59,090 --> 00:19:01,880
And so you can think about
that as refining the rule,

427
00:19:01,880 --> 00:19:03,860
that's like the learning side.

428
00:19:03,860 --> 00:19:07,700
Applying the rule, that's the
memory or decision making side.

429
00:19:07,700 --> 00:19:10,690
So you can think about,
during quiet wakefulness,

430
00:19:10,690 --> 00:19:14,870
this reactivation being used in
the service of actual learning

431
00:19:14,870 --> 00:19:18,090
or in decision making.

432
00:19:18,090 --> 00:19:20,540
And so again, you
read the paper,

433
00:19:20,540 --> 00:19:23,000
and one of the things that we
had discovered interestingly

434
00:19:23,000 --> 00:19:25,910
about reactivation
during quiet wakefulness

435
00:19:25,910 --> 00:19:28,610
was that when animals stop
after running on a track,

436
00:19:28,610 --> 00:19:32,230
they do reactivate sequences.

437
00:19:32,230 --> 00:19:35,090
This is raw data, animal
running from left to right

438
00:19:35,090 --> 00:19:36,929
and then stopping for
a long period of time.

439
00:19:36,929 --> 00:19:37,970
And you can see activity.

440
00:19:37,970 --> 00:19:39,560
There are these
bursts of activity.

441
00:19:39,560 --> 00:19:42,101
You blow these things up, these
are these sharp wave ripples.

442
00:19:42,101 --> 00:19:43,820
They last about
half a second or so.

443
00:19:43,820 --> 00:19:45,560
So you see a burst
of activity, you

444
00:19:45,560 --> 00:19:46,964
see these place cell sequences.

445
00:19:46,964 --> 00:19:48,380
In this case the
sequence actually

446
00:19:48,380 --> 00:19:50,940
runs in time reversed fashion.

447
00:19:50,940 --> 00:19:53,000
So in this case--

448
00:19:53,000 --> 00:19:57,770
time reversed fashion-- this
doesn't seem like planning

449
00:19:57,770 --> 00:19:59,330
or decision making.

450
00:19:59,330 --> 00:20:03,230
It may be evaluation of
temporal correlations

451
00:20:03,230 --> 00:20:06,236
that might be relevant
to behavior and learning.

452
00:20:06,236 --> 00:20:07,610
But what would a
reverse sequence

453
00:20:07,610 --> 00:20:12,500
actually have to do with
spatial learning and memory?

454
00:20:16,470 --> 00:20:18,990
The insight into how
this might be used

455
00:20:18,990 --> 00:20:20,820
came from computational models.

456
00:20:20,820 --> 00:20:23,400
In fact, computational models
that the post-doc in this case,

457
00:20:23,400 --> 00:20:26,280
Dave Foster, who
made this discovery,

458
00:20:26,280 --> 00:20:28,297
had used in his doctoral work.

459
00:20:28,297 --> 00:20:29,880
Where he was actually
building models,

460
00:20:29,880 --> 00:20:32,670
reinforcement learning
models of spatial navigation.

461
00:20:32,670 --> 00:20:35,430
And one of the problems
in reinforcement learning

462
00:20:35,430 --> 00:20:37,560
is that reinforcement
often comes

463
00:20:37,560 --> 00:20:41,040
after you've actually
carried out the steps that

464
00:20:41,040 --> 00:20:42,103
lead up to it.

465
00:20:42,103 --> 00:20:44,220
In other words, you
walk from left to right,

466
00:20:44,220 --> 00:20:46,044
you get rewarded
when you get here.

467
00:20:46,044 --> 00:20:47,460
What you want to
know is not just,

468
00:20:47,460 --> 00:20:48,630
this is where the reward is.

469
00:20:48,630 --> 00:20:50,604
What you really want
to know is, what

470
00:20:50,604 --> 00:20:52,520
were the things that
actually lead up to that?

471
00:20:52,520 --> 00:20:56,310
In other words, I want to
take credit, reward value,

472
00:20:56,310 --> 00:20:59,680
and I want to spread
it backward in time

473
00:20:59,680 --> 00:21:02,490
to place value on the
things that predict

474
00:21:02,490 --> 00:21:05,110
or lead up to reward.

475
00:21:05,110 --> 00:21:07,259
The so-called temporal
credit assignment problem.

476
00:21:07,259 --> 00:21:09,550
How do I give credit to things
that actually lead up to

477
00:21:09,550 --> 00:21:11,040
or predict reward?

478
00:21:11,040 --> 00:21:14,490
And thinking about how you
might solve the temporal credit

479
00:21:14,490 --> 00:21:17,580
assignment problem, this
reverse reactivation

480
00:21:17,580 --> 00:21:22,060
actually has the capacity to
solve that problem in one step.

481
00:21:22,060 --> 00:21:24,930
If you imagine when the
animal gets to the end,

482
00:21:24,930 --> 00:21:27,000
gets some reward,
and if now at that

483
00:21:27,000 --> 00:21:30,060
point I pair the delivery
of reward signal,

484
00:21:30,060 --> 00:21:31,590
which I will indicate here as--

485
00:21:31,590 --> 00:21:34,040
this is a cartoon
suggesting this is dopamine,

486
00:21:34,040 --> 00:21:35,610
a reward signal.

487
00:21:35,610 --> 00:21:39,910
And I'm going to pair that
with the reverse sequence.

488
00:21:39,910 --> 00:21:42,210
And now you can think
of the association.

489
00:21:42,210 --> 00:21:43,410
This is my current location.

490
00:21:43,410 --> 00:21:45,930
This is way back
where I started from.

491
00:21:45,930 --> 00:21:50,970
Current location gets
paired with strong reward.

492
00:21:50,970 --> 00:21:55,580
Remote location gets
paired with low reward.

493
00:21:55,580 --> 00:21:59,630
So this association
will essentially solve,

494
00:21:59,630 --> 00:22:05,270
will convolve this discrete
reward impulse function,

495
00:22:05,270 --> 00:22:09,950
will turn it into
this continuous graded

496
00:22:09,950 --> 00:22:13,130
monotonic reward
gradient function.

497
00:22:13,130 --> 00:22:15,860
Translating this into this
essentially in one step.

498
00:22:15,860 --> 00:22:17,870
So the thinking is, hey,
animals are actually

499
00:22:17,870 --> 00:22:19,790
using this to learn,
and they're using

500
00:22:19,790 --> 00:22:22,190
this to solve the
temporal credit assignment

501
00:22:22,190 --> 00:22:25,460
problem in a way that
would be important for

502
00:22:25,460 --> 00:22:27,800
general reinforcement learning.

503
00:22:27,800 --> 00:22:29,342
Now, I won't go into--

504
00:22:29,342 --> 00:22:30,800
so, we actually
did this experiment

505
00:22:30,800 --> 00:22:36,210
recorded from the reward area,
the VTA, and the hippocampus.

506
00:22:36,210 --> 00:22:40,580
And, indeed, during these
reactivation events,

507
00:22:40,580 --> 00:22:46,100
you see the pairing of
this reward signaling.

508
00:22:46,100 --> 00:22:48,680
And not only do
you see the pairing

509
00:22:48,680 --> 00:22:50,660
of the reward
signaling, but you find

510
00:22:50,660 --> 00:22:52,610
pairing of the reward
signaling in which

511
00:22:52,610 --> 00:22:57,610
the precise firing
of reward signals

512
00:22:57,610 --> 00:23:02,200
map onto the delivery of
rewards at goal locations.

513
00:23:02,200 --> 00:23:06,980
So it's not just certain
sequences are good,

514
00:23:06,980 --> 00:23:08,210
others are bad.

515
00:23:08,210 --> 00:23:09,800
It's that there are
certain locations

516
00:23:09,800 --> 00:23:12,620
along the sequence that have
differential reward value.

517
00:23:12,620 --> 00:23:16,020
So there's a mapping of
relative reward to location.

518
00:23:16,020 --> 00:23:21,050
And that if you look at this
while animals are actually

519
00:23:21,050 --> 00:23:25,400
performing a task you can see
biases in these sequences that

520
00:23:25,400 --> 00:23:26,900
correspond to planning.

521
00:23:26,900 --> 00:23:29,217
So both of these
elements of sequence

522
00:23:29,217 --> 00:23:31,467
reactivation-- this was some
work by Dave Foster where

523
00:23:31,467 --> 00:23:33,758
he looked at an animal that
has to just forage and find

524
00:23:33,758 --> 00:23:35,270
a location space.

525
00:23:35,270 --> 00:23:37,117
When animals are
searching for a location,

526
00:23:37,117 --> 00:23:38,950
and you look at these
reactivated sequences,

527
00:23:38,950 --> 00:23:41,570
it turns out that
reactivated sequences

528
00:23:41,570 --> 00:23:44,410
tend to be directed
toward the locations

529
00:23:44,410 --> 00:23:46,160
where the animal thinks
the goal might be.

530
00:23:46,160 --> 00:23:48,470
So it's sort of
thinking about things

531
00:23:48,470 --> 00:23:49,760
that would lead to reward.

532
00:23:49,760 --> 00:23:53,240
So both sides of this kind
of sequential computation

533
00:23:53,240 --> 00:23:55,540
seem to be expressed
in the hippocampus.

534
00:23:55,540 --> 00:23:59,510
Co-expression with reward
during evaluation or learning.

535
00:23:59,510 --> 00:24:02,480
The expression during
reward directed behavior

536
00:24:02,480 --> 00:24:05,150
in the service of planning
and decision making.

537
00:24:05,150 --> 00:24:11,650
And then finally, this is
the paper that you had read.

538
00:24:11,650 --> 00:24:14,580
It's just looking at the
structure of these reward

539
00:24:14,580 --> 00:24:17,180
sequences on long tracks.

540
00:24:17,180 --> 00:24:22,700
And the most salient point
that came out of this paper

541
00:24:22,700 --> 00:24:26,930
is that the phenomena of
sharp wave ripple sequential

542
00:24:26,930 --> 00:24:30,100
activation that correspond
to these theta sequences,

543
00:24:30,100 --> 00:24:34,040
when animals are sitting quietly
over the course of these longer

544
00:24:34,040 --> 00:24:38,550
up state like events, takes the
form of bursts of sharp wave

545
00:24:38,550 --> 00:24:41,790
ripples, or sequences of
these short sequences.

546
00:24:41,790 --> 00:24:43,730
So there's this
compositional notion

547
00:24:43,730 --> 00:24:46,464
that you form long sequences
out of little short sequences.

548
00:24:46,464 --> 00:24:48,380
And the other question
is, how do you actually

549
00:24:48,380 --> 00:24:50,270
put these short
sequences together,

550
00:24:50,270 --> 00:24:55,700
and why would you have
a representation that

551
00:24:55,700 --> 00:24:57,730
has this kind of
compositional structure?

552
00:24:57,730 --> 00:24:59,690
It's the Lego
block kind of idea,

553
00:24:59,690 --> 00:25:04,700
where I can put together
sequences from these more

554
00:25:04,700 --> 00:25:09,440
elemental sequential units.

555
00:25:09,440 --> 00:25:11,630
I won't show the movie,
but if I actually

556
00:25:11,630 --> 00:25:13,256
zoom in on one of
those sequences

557
00:25:13,256 --> 00:25:15,380
in that particular instance
where the animals stop,

558
00:25:15,380 --> 00:25:17,900
you can see there's
a longer reactivated

559
00:25:17,900 --> 00:25:21,320
sequence that actually takes the
form of four shorter sequences.

560
00:25:21,320 --> 00:25:24,340
That's how long sequences
are being evaluated.

561
00:25:29,630 --> 00:25:31,130
And one interesting
thing about this

562
00:25:31,130 --> 00:25:33,171
was that if you look at
these different sequences

563
00:25:33,171 --> 00:25:37,580
of different length, shorter
sequences, longer sequences,

564
00:25:37,580 --> 00:25:39,430
they all seem to have
a fixed velocity.

565
00:25:39,430 --> 00:25:41,132
And that is, there's this kind
of a fixed-- what I've heard

566
00:25:41,132 --> 00:25:42,132
is the speed of thought.

567
00:25:42,132 --> 00:25:43,890
So there is some
fixed constraint

568
00:25:43,890 --> 00:25:48,920
that evaluating further
out in the future

569
00:25:48,920 --> 00:25:50,785
simply takes more time.

570
00:25:50,785 --> 00:25:52,160
And so that suggests
that there's

571
00:25:52,160 --> 00:25:56,090
an interesting
constraint in capacity

572
00:25:56,090 --> 00:25:58,670
for extended
evaluation that comes

573
00:25:58,670 --> 00:26:05,420
in the form of these oscillatory
modes of the slow oscillation.

574
00:26:05,420 --> 00:26:08,390
Longer, or slower, frequencies
give you, essentially,

575
00:26:08,390 --> 00:26:11,700
more time, longer sequences.

576
00:26:11,700 --> 00:26:14,547
And that, potentially, if
you want to go even longer,

577
00:26:14,547 --> 00:26:16,130
you might actually
even link sequences

578
00:26:16,130 --> 00:26:17,570
across subsequent cycles.

579
00:26:17,570 --> 00:26:20,264
Just as you could do across
successive sharp wave ripples,

580
00:26:20,264 --> 00:26:22,680
you might be able to do that
across successive slow waves.

581
00:26:22,680 --> 00:26:25,520
So the idea that you have these
oscillations that you can link

582
00:26:25,520 --> 00:26:28,309
or couple sequences
across different cycles

583
00:26:28,309 --> 00:26:30,350
of these oscillations,
that you have oscillations

584
00:26:30,350 --> 00:26:32,420
at different
frequencies, suggests

585
00:26:32,420 --> 00:26:34,670
that this is the mechanism.

586
00:26:34,670 --> 00:26:37,310
It's the compositional
mechanism for sequences that

587
00:26:37,310 --> 00:26:40,160
are nested within oscillations.

588
00:26:40,160 --> 00:26:42,500
By combining these
oscillations you

589
00:26:42,500 --> 00:26:45,950
can construct these sequences in
ways that presumably contribute

590
00:26:45,950 --> 00:26:47,850
to cognition.

591
00:26:47,850 --> 00:26:50,449
So we now kind of see how
things are structured,

592
00:26:50,449 --> 00:26:52,740
and now the question is, can
we actually manipulate it?

593
00:26:52,740 --> 00:26:55,000
And so that's really
the challenge.

594
00:26:55,000 --> 00:26:57,290
Seeing the
compositional structure,

595
00:26:57,290 --> 00:27:00,050
having demonstrated potential
access, the ability to bias

596
00:27:00,050 --> 00:27:00,830
content.

597
00:27:00,830 --> 00:27:06,150
To coord, to bias the
structure of when things occur,

598
00:27:06,150 --> 00:27:07,650
what actually occurs.

599
00:27:07,650 --> 00:27:12,680
Now I think the capacity
exists to essentially engineer

600
00:27:12,680 --> 00:27:13,970
the sequences themselves.

601
00:27:13,970 --> 00:27:15,980
And that is that, during the
sleep or quiet wakeful states,

602
00:27:15,980 --> 00:27:18,230
get animals to think about
things that they have not

603
00:27:18,230 --> 00:27:22,280
actually experienced, and
test the hypothesis that this

604
00:27:22,280 --> 00:27:25,490
is what animals are using
to construct these models,

605
00:27:25,490 --> 00:27:29,660
to essentially tinker with the
blocks of memory and cognition.

606
00:27:29,660 --> 00:27:33,170
And then finally, just
pointing out, it's interesting.

607
00:27:33,170 --> 00:27:35,885
These ripple sequences,
same space and time scale

608
00:27:35,885 --> 00:27:36,909
as theta sequences.

609
00:27:36,909 --> 00:27:39,200
So again, it's suggesting
this is the fundamental unit.

610
00:27:39,200 --> 00:27:41,240
It's not unique to
the hippocampus.

611
00:27:41,240 --> 00:27:43,130
Essentially any
brain system could

612
00:27:43,130 --> 00:27:47,420
express this kind of structure
using that simple model.

613
00:27:47,420 --> 00:27:49,400
Where you see ramps and
you see oscillations,

614
00:27:49,400 --> 00:27:51,000
you're going to get sequences.

615
00:27:51,000 --> 00:27:54,740
And this capacity is something
that probably is broadly

616
00:27:54,740 --> 00:27:56,910
expressed and enjoyed by
different brain areas.

617
00:27:56,910 --> 00:28:00,100
So, there you go.