[Script Info]
Title: 
[Events]
Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text
Dialogue: 0,0:00:01.33,0:00:05.15,Default,,0000,0000,0000,,Computer algorithms today\Nare performing incredible tasks
Dialogue: 0,0:00:05.17,0:00:09.68,Default,,0000,0000,0000,,with high accuracies, at a massive scale,\Nusing human-like intelligence.
Dialogue: 0,0:00:10.05,0:00:13.86,Default,,0000,0000,0000,,And this intelligence of computers\Nis often referred to as AI
Dialogue: 0,0:00:13.89,0:00:15.74,Default,,0000,0000,0000,,or artificial intelligence.
Dialogue: 0,0:00:15.76,0:00:20.43,Default,,0000,0000,0000,,AI is poised to make an incredible impact\Non our lives in the future.
Dialogue: 0,0:00:20.91,0:00:24.81,Default,,0000,0000,0000,,Today, however, we still\Nface massive challenges
Dialogue: 0,0:00:24.83,0:00:28.32,Default,,0000,0000,0000,,in detecting and diagnosing\Nseveral life-threatening illnesses,
Dialogue: 0,0:00:28.35,0:00:30.85,Default,,0000,0000,0000,,such as infectious diseases and cancer.
Dialogue: 0,0:00:32.02,0:00:34.28,Default,,0000,0000,0000,,Thousands of patients, every year,
Dialogue: 0,0:00:34.30,0:00:37.43,Default,,0000,0000,0000,,lose their lives due to\Nliver and oral cancer.
Dialogue: 0,0:00:37.89,0:00:40.58,Default,,0000,0000,0000,,Our best way to help these patients
Dialogue: 0,0:00:40.60,0:00:44.91,Default,,0000,0000,0000,,is to perform early detection\Nand diagnoses of these diseases.
Dialogue: 0,0:00:45.87,0:00:50.26,Default,,0000,0000,0000,,So how do we detect these diseases today\Nand can artificial intelligence help?
Dialogue: 0,0:00:51.97,0:00:55.59,Default,,0000,0000,0000,,In patients who, unfortunately,\Nare suspected of these diseases,
Dialogue: 0,0:00:55.62,0:00:58.24,Default,,0000,0000,0000,,an expert physician first orders
Dialogue: 0,0:00:58.27,0:01:00.89,Default,,0000,0000,0000,,very expensive medical\Nimaging technologies
Dialogue: 0,0:01:00.92,0:01:04.67,Default,,0000,0000,0000,,such as fluorescent imaging, CTs,\NMRIs, to be performed.
Dialogue: 0,0:01:05.05,0:01:07.34,Default,,0000,0000,0000,,Once those images are collected,
Dialogue: 0,0:01:07.36,0:01:12.11,Default,,0000,0000,0000,,another expert physician then diagnoses\Nthose images and talks to the patient.
Dialogue: 0,0:01:12.52,0:01:15.96,Default,,0000,0000,0000,,As you can see, this is a very\Nresource-intensive process,
Dialogue: 0,0:01:15.98,0:01:20.42,Default,,0000,0000,0000,,requiring both expert physicians,\Nexpensive medical imaging technologies,
Dialogue: 0,0:01:20.44,0:01:23.52,Default,,0000,0000,0000,,and is not considered practical\Nfor the developing world.
Dialogue: 0,0:01:23.54,0:01:27.04,Default,,0000,0000,0000,,And in fact, in many\Nindustrialized nations, as well.
Dialogue: 0,0:01:27.78,0:01:31.07,Default,,0000,0000,0000,,So, can we solve this problem\Nusing artificial intelligence?
Dialogue: 0,0:01:31.89,0:01:35.90,Default,,0000,0000,0000,,Today, if I were to use traditional\Nartificial intelligence architectures
Dialogue: 0,0:01:35.93,0:01:37.13,Default,,0000,0000,0000,,to solve this problem,
Dialogue: 0,0:01:37.16,0:01:38.74,Default,,0000,0000,0000,,I would require 10,000 --
Dialogue: 0,0:01:38.76,0:01:42.66,Default,,0000,0000,0000,,I repeat, on an order of 10,000\Nof these very expensive medical images
Dialogue: 0,0:01:42.69,0:01:44.08,Default,,0000,0000,0000,,first to be generated.
Dialogue: 0,0:01:44.11,0:01:46.97,Default,,0000,0000,0000,,After that, I would then go\Nto an expert physician,
Dialogue: 0,0:01:46.100,0:01:49.51,Default,,0000,0000,0000,,who would then analyze\Nthose images for me.
Dialogue: 0,0:01:49.54,0:01:51.70,Default,,0000,0000,0000,,And using those two pieces of information,
Dialogue: 0,0:01:51.72,0:01:55.30,Default,,0000,0000,0000,,I can train a standard deep neural network\Nor a deep learning network,
Dialogue: 0,0:01:55.32,0:01:57.46,Default,,0000,0000,0000,,to provide patient's diagnosis.
Dialogue: 0,0:01:57.48,0:01:59.20,Default,,0000,0000,0000,,Similar to the first approach,
Dialogue: 0,0:01:59.22,0:02:01.39,Default,,0000,0000,0000,,traditional artificial\Nintelligence approaches
Dialogue: 0,0:02:01.41,0:02:02.86,Default,,0000,0000,0000,,suffer from the same problem.
Dialogue: 0,0:02:02.88,0:02:07.32,Default,,0000,0000,0000,,Large amounts of data, expert physicians,\Nand expert medical imaging technologies.
Dialogue: 0,0:02:08.35,0:02:12.60,Default,,0000,0000,0000,,So can we invent more scalable, effective,
Dialogue: 0,0:02:12.62,0:02:15.93,Default,,0000,0000,0000,,and more valuable\Nartificial intelligence architectures
Dialogue: 0,0:02:15.96,0:02:18.69,Default,,0000,0000,0000,,to solve these very important\Nproblems facing us today?
Dialogue: 0,0:02:19.03,0:02:22.05,Default,,0000,0000,0000,,And this is exactly what my group\Nat MIT Media Lab does.
Dialogue: 0,0:02:22.38,0:02:26.20,Default,,0000,0000,0000,,We have invented a variety\Nof unorthodox AI architectures
Dialogue: 0,0:02:26.22,0:02:29.43,Default,,0000,0000,0000,,to solve some of the most important\Nchallenges facing us today
Dialogue: 0,0:02:29.45,0:02:31.90,Default,,0000,0000,0000,,in medical imaging and clinical trials.
Dialogue: 0,0:02:32.48,0:02:35.54,Default,,0000,0000,0000,,In the example I shared\Nwith you today, we had two goals.
Dialogue: 0,0:02:35.56,0:02:38.52,Default,,0000,0000,0000,,Our first goal was to reduce\Nthe number of images
Dialogue: 0,0:02:38.55,0:02:41.82,Default,,0000,0000,0000,,required to train artificial\Nintelligence algorithms.
Dialogue: 0,0:02:41.84,0:02:43.93,Default,,0000,0000,0000,,Our second goal -- we were more ambitious,
Dialogue: 0,0:02:43.95,0:02:47.71,Default,,0000,0000,0000,,we wanted to reduce the use\Nof expensive medical imaging technologies
Dialogue: 0,0:02:47.74,0:02:48.93,Default,,0000,0000,0000,,to screen patients.
Dialogue: 0,0:02:48.96,0:02:50.29,Default,,0000,0000,0000,,So how did we do it?
Dialogue: 0,0:02:50.97,0:02:52.15,Default,,0000,0000,0000,,For our first goal,
Dialogue: 0,0:02:52.17,0:02:54.22,Default,,0000,0000,0000,,instead of starting\Nwith tens and thousands
Dialogue: 0,0:02:54.24,0:02:57.25,Default,,0000,0000,0000,,of these very expensive medical images,\Nlike traditional AI,
Dialogue: 0,0:02:57.28,0:02:59.36,Default,,0000,0000,0000,,we started with a single medical image.
Dialogue: 0,0:02:59.38,0:03:03.12,Default,,0000,0000,0000,,From this image, my team and I\Nfigured out a very clever way
Dialogue: 0,0:03:03.15,0:03:05.90,Default,,0000,0000,0000,,to extract billions\Nof information packets.
Dialogue: 0,0:03:05.92,0:03:09.63,Default,,0000,0000,0000,,These information packets\Nincluded colors, pixels, geometry
Dialogue: 0,0:03:09.65,0:03:12.18,Default,,0000,0000,0000,,and rendering of the disease\Non the medical image.
Dialogue: 0,0:03:12.20,0:03:16.54,Default,,0000,0000,0000,,In a sense, we converted one image\Ninto billions of training data points,
Dialogue: 0,0:03:16.56,0:03:19.54,Default,,0000,0000,0000,,massively reducing the amount of data\Nneeded for training.
Dialogue: 0,0:03:20.15,0:03:21.32,Default,,0000,0000,0000,,For our second goal,
Dialogue: 0,0:03:21.35,0:03:25.22,Default,,0000,0000,0000,,to reduce the use of expensive medical\Nimaging technologies to screen patients,
Dialogue: 0,0:03:25.24,0:03:28.09,Default,,0000,0000,0000,,we started with a standard,\Nwhite light photograph,
Dialogue: 0,0:03:28.12,0:03:32.48,Default,,0000,0000,0000,,acquired either from a DSLR camera\Nor a mobile phone, for the patient.
Dialogue: 0,0:03:32.51,0:03:34.97,Default,,0000,0000,0000,,Then remember those\Nbillions of information packets?
Dialogue: 0,0:03:34.100,0:03:38.50,Default,,0000,0000,0000,,We overlaid those from\Nthe medical image onto this image,
Dialogue: 0,0:03:38.52,0:03:41.47,Default,,0000,0000,0000,,creating something\Nwhat we call a composite image.
Dialogue: 0,0:03:41.50,0:03:44.78,Default,,0000,0000,0000,,Much to our surprise,\Nwe only required 50 --
Dialogue: 0,0:03:44.80,0:03:46.13,Default,,0000,0000,0000,,I repeat, only 50 --
Dialogue: 0,0:03:46.16,0:03:50.02,Default,,0000,0000,0000,,of these composite images to train\Nour algorithms to high efficiencies.
Dialogue: 0,0:03:50.70,0:03:52.03,Default,,0000,0000,0000,,To summarize our approach,
Dialogue: 0,0:03:52.05,0:03:55.22,Default,,0000,0000,0000,,instead of using 10,000\Nvery expensive medical images,
Dialogue: 0,0:03:55.24,0:03:58.25,Default,,0000,0000,0000,,we can now train the AI algorithms\Nin an unorthodox way,
Dialogue: 0,0:03:58.28,0:04:02.53,Default,,0000,0000,0000,,using only 50 of these high-resolution,\Nbut standard photographs,
Dialogue: 0,0:04:02.55,0:04:05.07,Default,,0000,0000,0000,,acquired from DSLR cameras\Nand mobile phones
Dialogue: 0,0:04:05.09,0:04:06.62,Default,,0000,0000,0000,,and provide diagnosis.
Dialogue: 0,0:04:06.64,0:04:07.81,Default,,0000,0000,0000,,More importantly,
Dialogue: 0,0:04:07.83,0:04:11.03,Default,,0000,0000,0000,,our algorithms can accept,\Nin the future and even right now,
Dialogue: 0,0:04:11.05,0:04:13.84,Default,,0000,0000,0000,,some very simple, white light\Nphotographs from the patient,
Dialogue: 0,0:04:13.86,0:04:16.59,Default,,0000,0000,0000,,instead of expensive medical\Nimaging technologies.
Dialogue: 0,0:04:17.16,0:04:20.23,Default,,0000,0000,0000,,I believe that we are poised\Nto enter an era
Dialogue: 0,0:04:20.25,0:04:22.17,Default,,0000,0000,0000,,where artificial intelligence
Dialogue: 0,0:04:22.19,0:04:24.72,Default,,0000,0000,0000,,is going to make an incredible\Nimpact on our future.
Dialogue: 0,0:04:24.74,0:04:27.20,Default,,0000,0000,0000,,And I think that thinking\Nabout traditional AI,
Dialogue: 0,0:04:27.23,0:04:29.100,Default,,0000,0000,0000,,which is data-rich but application-poor,
Dialogue: 0,0:04:30.02,0:04:31.57,Default,,0000,0000,0000,,we should also continue thinking
Dialogue: 0,0:04:31.59,0:04:34.60,Default,,0000,0000,0000,,about unorthodox artificial\Nintelligence architectures,
Dialogue: 0,0:04:34.62,0:04:36.57,Default,,0000,0000,0000,,which can accept small amounts of data
Dialogue: 0,0:04:36.59,0:04:39.54,Default,,0000,0000,0000,,and solve some of the most important\Nproblems facing us today,
Dialogue: 0,0:04:39.56,0:04:40.80,Default,,0000,0000,0000,,especially in healthcare.
Dialogue: 0,0:04:40.82,0:04:42.00,Default,,0000,0000,0000,,Thank you very much.
Dialogue: 0,0:04:42.02,0:04:45.83,Default,,0000,0000,0000,,(Applause)