0:00:01.280,0:00:05.136
Computer algorithms today[br]are performing incredible tasks

0:00:05.160,0:00:09.896
with high accuracies, at a massive scale,[br]using human-like intelligence.

0:00:09.920,0:00:13.856
And this intelligence of computers[br]is often referred to as AI

0:00:13.880,0:00:15.736
or artificial intelligence.

0:00:15.760,0:00:19.960
AI is poised to make an incredible impact[br]on our lives in the future.

0:00:20.880,0:00:24.816
Today, however,[br]we still face massive challenges

0:00:24.840,0:00:28.336
in detecting and diagnosing[br]several life-threatening illnesses,

0:00:28.360,0:00:30.720
such as infectious diseases and cancer.

0:00:32.000,0:00:34.296
Thousands of patients every year

0:00:34.320,0:00:37.120
lose their lives[br]due to liver and oral cancer.

0:00:37.880,0:00:40.576
Our best way to help these patients

0:00:40.600,0:00:44.920
is to perform early detection[br]and diagnoses of these diseases.

0:00:45.880,0:00:50.040
So how do we detect these diseases today,[br]and can artificial intelligence help?

0:00:51.920,0:00:55.576
In patients who, unfortunately,[br]are suspected of these diseases,

0:00:55.600,0:00:58.256
an expert physician first orders

0:00:58.280,0:01:00.896
very expensive[br]medical imaging technologies

0:01:00.920,0:01:05.016
such as fluorescent imaging,[br]CTs, MRIs, to be performed.

0:01:05.040,0:01:07.336
Once those images are collected,

0:01:07.360,0:01:11.880
another expert physician then diagnoses[br]those images and talks to the patient.

0:01:12.520,0:01:15.976
As you can see, this is[br]a very resource-intensive process,

0:01:16.000,0:01:20.416
requiring both expert physicians,[br]expensive medical imaging technologies,

0:01:20.440,0:01:23.536
and is not considered practical[br]for the developing world.

0:01:23.560,0:01:26.920
And in fact, in many[br]industrialized nations, as well.

0:01:27.760,0:01:30.640
So, can we solve this problem[br]using artificial intelligence?

0:01:31.840,0:01:35.896
Today, if I were to use traditional[br]artificial intelligence architectures

0:01:35.920,0:01:37.136
to solve this problem,

0:01:37.160,0:01:38.616
I would require 10,000 --

0:01:38.640,0:01:42.656
I repeat, on an order of 10,000[br]of these very expensive medical images

0:01:42.680,0:01:44.056
first to be generated.

0:01:44.080,0:01:46.976
After that, I would then go[br]to an expert physician,

0:01:47.000,0:01:49.496
who would then analyze[br]those images for me.

0:01:49.520,0:01:51.616
And using those two pieces of information,

0:01:51.640,0:01:55.296
I can train a standard deep neural network[br]or a deep learning network

0:01:55.320,0:01:57.456
to provide patient's diagnosis.

0:01:57.480,0:01:59.216
Similar to the first approach,

0:01:59.240,0:02:01.383
traditional artificial[br]intelligence approaches

0:02:01.407,0:02:02.856
suffer from the same problem.

0:02:02.880,0:02:07.440
Large amounts of data, expert physicians[br]and expert medical imaging technologies.

0:02:08.320,0:02:12.616
So, can we invent more scalable, effective

0:02:12.640,0:02:15.936
and more valuable artificial[br]intelligence architectures

0:02:15.960,0:02:19.016
to solve these very important[br]problems facing us today?

0:02:19.040,0:02:22.336
And this is exactly[br]what my group at MIT Media Lab does.

0:02:22.360,0:02:26.216
We have invented a variety[br]of unorthodox AI architectures

0:02:26.240,0:02:29.416
to solve some of the most important[br]challenges facing us today

0:02:29.440,0:02:31.640
in medical imaging and clinical trials.

0:02:32.480,0:02:35.536
In the example I shared[br]with you today, we had two goals.

0:02:35.560,0:02:38.536
Our first goal was to reduce[br]the number of images

0:02:38.560,0:02:41.816
required to train[br]artificial intelligence algorithms.

0:02:41.840,0:02:43.936
Our second goal -- we were more ambitious,

0:02:43.960,0:02:47.696
we wanted to reduce the use[br]of expensive medical imaging technologies

0:02:47.720,0:02:48.936
to screen patients.

0:02:48.960,0:02:50.160
So how did we do it?

0:02:50.920,0:02:52.136
For our first goal,

0:02:52.160,0:02:54.216
instead of starting[br]with tens and thousands

0:02:54.240,0:02:57.256
of these very expensive medical images,[br]like traditional AI,

0:02:57.280,0:02:59.336
we started with a single medical image.

0:02:59.360,0:03:03.136
From this image, my team and I[br]figured out a very clever way

0:03:03.160,0:03:05.896
to extract billions[br]of information packets.

0:03:05.920,0:03:09.616
These information packets[br]included colors, pixels, geometry

0:03:09.640,0:03:12.176
and rendering of the disease[br]on the medical image.

0:03:12.200,0:03:16.536
In a sense, we converted one image[br]into billions of training data points,

0:03:16.560,0:03:20.096
massively reducing the amount of data[br]needed for training.

0:03:20.120,0:03:21.336
For our second goal,

0:03:21.360,0:03:25.216
to reduce the use of expensive medical[br]imaging technologies to screen patients,

0:03:25.240,0:03:28.096
we started with a standard,[br]white light photograph,

0:03:28.120,0:03:32.456
acquired either from a DSLR camera[br]or a mobile phone, for the patient.

0:03:32.480,0:03:34.936
Then remember those[br]billions of information packets?

0:03:34.960,0:03:38.496
We overlaid those from[br]the medical image onto this image,

0:03:38.520,0:03:41.040
creating something[br]that we call a composite image.

0:03:41.480,0:03:44.776
Much to our surprise,[br]we only required 50 --

0:03:44.800,0:03:46.136
I repeat, only 50 --

0:03:46.160,0:03:50.000
of these composite images to train[br]our algorithms to high efficiencies.

0:03:50.680,0:03:52.016
To summarize our approach,

0:03:52.040,0:03:55.216
instead of using 10,000[br]very expensive medical images,

0:03:55.240,0:03:58.256
we can now train the AI algorithms[br]in an unorthodox way,

0:03:58.280,0:04:02.536
using only 50 of these high-resolution,[br]but standard photographs,

0:04:02.560,0:04:05.056
acquired from DSLR cameras[br]and mobile phones,

0:04:05.080,0:04:06.616
and provide diagnosis.

0:04:06.640,0:04:07.856
More importantly,

0:04:07.880,0:04:11.016
our algorithms can accept,[br]in the future and even right now,

0:04:11.040,0:04:13.856
some very simple, white light[br]photographs from the patient,

0:04:13.880,0:04:16.320
instead of expensive[br]medical imaging technologies.

0:04:17.120,0:04:20.216
I believe that we are poised[br]to enter an era

0:04:20.240,0:04:22.176
where artificial intelligence

0:04:22.200,0:04:24.736
is going to make an incredible[br]impact on our future.

0:04:24.760,0:04:27.216
And I think that thinking[br]about traditional AI,

0:04:27.240,0:04:30.016
which is data-rich but application-poor,

0:04:30.040,0:04:31.576
we should also continue thinking

0:04:31.600,0:04:34.616
about unorthodox artificial[br]intelligence architectures,

0:04:34.640,0:04:36.576
which can accept small amounts of data

0:04:36.600,0:04:39.536
and solve some of the most important[br]problems facing us today,

0:04:39.560,0:04:40.816
especially in health care.

0:04:40.840,0:04:42.056
Thank you very much.

0:04:42.080,0:04:45.920
(Applause)