WEBVTT

00:00:05.302 --> 00:00:08.194
[MUSIC PLAYING]

00:00:11.443 --> 00:00:13.110
AMIT GANDHI: Hi, my
name is Amit Gandhi,

00:00:13.110 --> 00:00:15.180
and I'm a graduate
researcher at MIT.

00:00:15.180 --> 00:00:17.430
Welcome to this course
on exploring fairness

00:00:17.430 --> 00:00:19.890
in machine learning for
international development.

00:00:19.890 --> 00:00:22.350
I'm going to present the
motivation for this course, why

00:00:22.350 --> 00:00:25.230
it is important to pay attention
to ethics and appropriate use

00:00:25.230 --> 00:00:28.080
in the topics we
will be covering.

00:00:28.080 --> 00:00:29.730
Let's start with an
introduction to how

00:00:29.730 --> 00:00:32.680
machine learning is being used
in international development.

00:00:32.680 --> 00:00:34.410
As a brief review,
machine learning

00:00:34.410 --> 00:00:36.443
is a branch of
artificial intelligence

00:00:36.443 --> 00:00:38.610
in which software learns
how to perform a task based

00:00:38.610 --> 00:00:42.570
on experience as opposed to
being explicitly programmed.

00:00:42.570 --> 00:00:44.640
As a result, this
emerging technology

00:00:44.640 --> 00:00:47.070
is widely applicable
across many fields

00:00:47.070 --> 00:00:48.930
and can leverage
the power of data

00:00:48.930 --> 00:00:51.670
in ways that were
previously impossible.

00:00:51.670 --> 00:00:53.860
Some application areas
for machine learning

00:00:53.860 --> 00:00:56.200
include medicine,
workforce development,

00:00:56.200 --> 00:00:58.180
and financial inclusion.

00:00:58.180 --> 00:01:00.130
In the next few slides,
I will talk about some

00:01:00.130 --> 00:01:02.110
of the ways machine
learning is being used,

00:01:02.110 --> 00:01:04.750
and we will go more in depth
about ethical usage of machine

00:01:04.750 --> 00:01:07.820
learning in later modules.

00:01:07.820 --> 00:01:10.190
In the health care sector,
significant advances

00:01:10.190 --> 00:01:12.950
in machine learning have
allowed for rapid diagnostics

00:01:12.950 --> 00:01:14.540
of medical conditions.

00:01:14.540 --> 00:01:16.790
Several organizations
are developing tools,

00:01:16.790 --> 00:01:19.010
ranging from sensors
to smartphone apps,

00:01:19.010 --> 00:01:21.860
for community health workers
to collect patient data,

00:01:21.860 --> 00:01:24.290
get novel insights into
patterns of disease spread

00:01:24.290 --> 00:01:26.570
and occurrence, and
perform diagnostics away

00:01:26.570 --> 00:01:29.600
from hospitals and clinics,
significantly increasing

00:01:29.600 --> 00:01:32.060
the access and quality
of medical care

00:01:32.060 --> 00:01:36.030
available to individuals
in remote areas.

00:01:36.030 --> 00:01:38.040
In workforce development,
machine learning

00:01:38.040 --> 00:01:41.220
can reduce unemployment rates
by pairing skilled individuals

00:01:41.220 --> 00:01:42.810
with appropriate jobs.

00:01:42.810 --> 00:01:45.330
For example, in
India, hiring managers

00:01:45.330 --> 00:01:47.850
pay more attention to
credentials than skills,

00:01:47.850 --> 00:01:49.920
making it difficult for
uncredentialed people

00:01:49.920 --> 00:01:52.080
to get hired or promoted.

00:01:52.080 --> 00:01:53.970
Aspiring Minds is
an Indian company

00:01:53.970 --> 00:01:55.650
that has developed
a computer test that

00:01:55.650 --> 00:01:57.990
determines applicants'
strengths and connects them

00:01:57.990 --> 00:02:00.510
to better paying jobs.

00:02:00.510 --> 00:02:03.265
In financial inclusion,
several organizations

00:02:03.265 --> 00:02:04.890
are using machine
learning to determine

00:02:04.890 --> 00:02:07.770
credit worthiness of
individuals in areas where

00:02:07.770 --> 00:02:09.780
other formalized
credit systems may not

00:02:09.780 --> 00:02:11.640
be available or accessible.

00:02:11.640 --> 00:02:14.520
These companies are deploying
pay-as-you-go services,

00:02:14.520 --> 00:02:17.850
ranging from solar lighting
systems to agricultural inputs,

00:02:17.850 --> 00:02:20.850
and are generating and gathering
non-traditional data on user

00:02:20.850 --> 00:02:22.740
assets and repayments.

00:02:22.740 --> 00:02:25.080
This data can be used to
determine credit worthiness

00:02:25.080 --> 00:02:28.440
using machine learning, enabling
individuals to access loans

00:02:28.440 --> 00:02:32.310
or financing that were
otherwise inaccessible.

00:02:32.310 --> 00:02:34.650
Machine learning techniques
have been around for decades

00:02:34.650 --> 00:02:37.170
but have increased in
usage in recent years.

00:02:37.170 --> 00:02:38.820
Over the past decade,
this technology

00:02:38.820 --> 00:02:40.272
has become more accessible.

00:02:40.272 --> 00:02:42.480
Machine learning is being
taught at many universities

00:02:42.480 --> 00:02:45.180
across the world, and several
data analysis platforms

00:02:45.180 --> 00:02:47.970
have released machine learning
libraries and toolboxes.

00:02:47.970 --> 00:02:51.570
For example, with Scikit-learn,
an open source Python library,

00:02:51.570 --> 00:02:52.998
someone familiar
with programming

00:02:52.998 --> 00:02:54.540
can get started with
training machine

00:02:54.540 --> 00:02:57.150
learning models
within a few days.

00:02:57.150 --> 00:02:59.550
While the applications of
machine learning are great,

00:02:59.550 --> 00:03:02.310
it is also important to
acknowledge its limitations.

00:03:02.310 --> 00:03:04.350
The impacts of machine
learning on society

00:03:04.350 --> 00:03:06.330
are still not well understood.

00:03:06.330 --> 00:03:08.370
Without careful attention
to these issues,

00:03:08.370 --> 00:03:10.020
we run the risk of
applying systems

00:03:10.020 --> 00:03:12.870
that are not only ineffective,
but could also harm people

00:03:12.870 --> 00:03:15.720
by reinforcing existing
patterns of social inequity

00:03:15.720 --> 00:03:17.727
and exclusion.

00:03:17.727 --> 00:03:19.310
There is a large
community of research

00:03:19.310 --> 00:03:20.990
in this area in
developed countries,

00:03:20.990 --> 00:03:23.420
such as the FAT
or AI Now groups.

00:03:23.420 --> 00:03:26.210
A well-known example is about
gender-differentiated credit

00:03:26.210 --> 00:03:27.230
scoring.

00:03:27.230 --> 00:03:29.090
Low-income women in
developing countries

00:03:29.090 --> 00:03:31.430
often lack credit histories
and formal income.

00:03:31.430 --> 00:03:34.040
Despite evidence showing
that these women default

00:03:34.040 --> 00:03:36.080
on loans less than
men, the lack of data

00:03:36.080 --> 00:03:37.940
makes it difficult for
lending organizations

00:03:37.940 --> 00:03:40.100
to provide these women loans.

00:03:40.100 --> 00:03:43.040
As a result, taking predatory
loans from informal lenders

00:03:43.040 --> 00:03:45.530
is a common practice.

00:03:45.530 --> 00:03:48.200
While many of these concepts
translate to ethics and machine

00:03:48.200 --> 00:03:49.850
learning, research
about fairness

00:03:49.850 --> 00:03:51.770
in machine learning
specifically for issues

00:03:51.770 --> 00:03:53.450
relevant to
international development

00:03:53.450 --> 00:03:55.010
is at an early stage.

00:03:55.010 --> 00:03:56.900
Ongoing efforts to
ensure ethical usage

00:03:56.900 --> 00:04:00.060
of this technology will be
discussed in the next video.

00:04:00.060 --> 00:04:02.660
We will cover content from an
appropriate usage framework

00:04:02.660 --> 00:04:04.910
developed by the Center
for digital development

00:04:04.910 --> 00:04:09.090
at the US Agency for
International Development.

00:04:09.090 --> 00:04:10.830
In the second set
of modules, we will

00:04:10.830 --> 00:04:12.782
present case studies
of organizations

00:04:12.782 --> 00:04:14.490
that are currently
using machine learning

00:04:14.490 --> 00:04:16.380
in international
development, and discuss

00:04:16.380 --> 00:04:19.868
potential ethical issues that
may arise in these examples.

00:04:19.868 --> 00:04:21.660
At a high level, we
will present approaches

00:04:21.660 --> 00:04:23.610
on how to address
these issues and what

00:04:23.610 --> 00:04:25.447
the outcomes could mean.

00:04:25.447 --> 00:04:27.780
In the third set of modules,
we will present a framework

00:04:27.780 --> 00:04:29.730
for addressing
ethical challenges.

00:04:29.730 --> 00:04:32.430
Content will include
protected attributes, fairness

00:04:32.430 --> 00:04:35.430
through unawareness, choices
and fairness criteria,

00:04:35.430 --> 00:04:38.270
and techniques for
mitigating bias.

00:04:38.270 --> 00:04:40.730
In the fourth set of modules,
we will present case studies

00:04:40.730 --> 00:04:43.760
from organizations with examples
of how machine learning could

00:04:43.760 --> 00:04:46.670
be implemented appropriately.

00:04:46.670 --> 00:04:48.910
Thank you for taking the
time to watch these videos,

00:04:48.910 --> 00:04:50.702
and we hope that you
will continue to watch

00:04:50.702 --> 00:04:53.660
the rest of the series.

00:04:53.660 --> 00:04:56.710
[MUSIC PLAYING]