MIT AGI: Artificial General Intelligence

00:00:00.000 | Welcome to course 6S099 Artificial General Intelligence

00:00:06.500 | We will explore the nature of intelligence

00:00:10.500 | From as much as possible an engineering perspective

00:00:14.500 | You will hear many voices

00:00:17.500 | My voice will be that of an engineer

00:00:21.500 | Our mission is to engineer intelligence

00:00:28.500 | The MIT motto is "Mind and Hand"

00:00:31.500 | What that means is we want to explore the fundamental science

00:00:38.500 | of what makes an intelligent system

00:00:42.500 | The core concepts behind our understanding of what is intelligence

00:00:48.500 | But we always want to ground it in the creation of intelligent systems

00:00:55.500 | We always want to be in the now, in today

00:00:59.500 | In understanding how today we can build artificial intelligence systems

00:01:03.500 | that can make for a better world

00:01:05.500 | That is the core for us here at MIT

00:01:09.500 | First and foremost, we're scientists and engineers

00:01:13.500 | Our goal is to engineer intelligence

00:01:16.500 | We want to provide with this approach a balance to

00:01:23.500 | the very important but over-represented view of artificial general intelligence

00:01:31.500 | The black box reasoning view

00:01:34.500 | Where the idea is once we know how to create a human level intelligence system

00:01:41.500 | How will society be impacted?

00:01:43.500 | Will robots take over and kill everyone?

00:01:50.500 | Will we achieve a utopia that will remove the need to do any of the messy jobs

00:01:55.500 | that will make us all extremely happy?

00:01:58.500 | Those kinds of beautiful philosophical concepts are interesting to explore

00:02:03.500 | But that's not what we're interested in doing

00:02:05.500 | I believe that from an engineering perspective

00:02:09.500 | We want to focus on the black box of AGI

00:02:13.500 | Start to build insights and intuitions about how we create systems

00:02:18.500 | that approach human level intelligence

00:02:21.500 | I believe we're very far away from creating

00:02:26.500 | anything resembling human level intelligence

00:02:30.500 | However, the dimension of the metric behind the word far

00:02:38.500 | may not be time

00:02:40.500 | In time, perhaps through a few breakthroughs

00:02:47.500 | Maybe even one breakthrough

00:02:49.500 | Everything can change

00:02:51.500 | But as we stand now

00:02:53.500 | Our current methods as we will explore from the various ideas and approaches

00:02:57.500 | and the guest speakers coming here over the next two weeks and beyond

00:03:02.500 | Our best understanding, our best intuition and insights

00:03:07.500 | are not yet at the level of reaching

00:03:12.500 | without a major leap and breakthrough

00:03:14.500 | and paradigm shift towards human level intelligence

00:03:17.500 | So it's not constructive to consider the impact of artificial intelligence

00:03:23.500 | To consider questions of safety and ethics

00:03:29.500 | Fundamental, extremely important questions

00:03:32.500 | It's not constructive to consider those questions

00:03:36.500 | without also deeply considering the black box

00:03:40.500 | of the actual methods of artificial intelligence

00:03:44.500 | Human level artificial intelligence

00:03:46.500 | And that's what I see, what I hope this course can be

00:03:50.500 | Its first iteration, its first exploratory attempt

00:03:56.500 | to try to look at different approaches of how we can engineer intelligence

00:04:01.500 | That's the role of MIT

00:04:03.500 | Its tradition of mind and hand

00:04:05.500 | It's to consider the big picture

00:04:07.500 | The future impact of society 10, 20, 30, 40 years out

00:04:11.500 | But fundamentally grounded in what kind of methods do we have today

00:04:16.500 | And what are their limitations and possibilities of achieving that

00:04:21.500 | The black box of AGI

00:04:23.500 | And in the future impact on society

00:04:29.500 | of creating artificial intelligence systems

00:04:32.500 | that get, become increasingly more intelligent

00:04:35.500 | The fundamental disagreement lies in the fact

00:04:40.500 | The very core of that black box

00:04:43.500 | Which is how hard is it to build an AGI system

00:04:48.500 | How hard is it to create a human level artificial intelligence system

00:04:52.500 | That's the open question for all of us

00:04:54.500 | From Josh Tenenbaum to Andrej Karpathy

00:05:00.500 | To folks from OpenAI to Boston Dynamics

00:05:04.500 | To brilliant leaders in various fields of artificial intelligence

00:05:08.500 | that will come here

00:05:09.500 | That's the open question

00:05:10.500 | How hard is it?

00:05:11.500 | There's been a lot of incredibly impressive results

00:05:15.500 | In deep learning, in neuroscience, in computational cognitive science

00:05:20.500 | In robotics

00:05:23.500 | But how far are we still to go to the AGI?

00:05:28.500 | That's the fundamental question that we need to explore

00:05:31.500 | Before we consider the questions, the future impact on society

00:05:38.500 | And the goal for this class is to build intuition

00:05:41.500 | One talk at a time, a project at a time

00:05:46.500 | Build intuition about where we stand

00:05:48.500 | About what the limitations of current approaches are

00:05:51.500 | How can we close the gap?

00:05:54.500 | A nice meme that I caught on Twitter recently

00:05:59.500 | Of the difference between the engineering approach

00:06:04.500 | At the very simplest of a Google intern

00:06:08.500 | Typing a for loop that just does a grid search

00:06:11.500 | On parameters for neural network

00:06:14.500 | And on the right is the way media would report this for loop

00:06:19.500 | The Google AI created its own baby AI

00:06:25.500 | I think it's easy for us to go one way or the other

00:06:31.500 | But we'd like to do both

00:06:33.500 | Our first goal is to avoid the pitfalls of black box thinking

00:06:38.500 | Of the futurism thinking that results in hype

00:06:41.500 | That's detached from scientific engineering understanding

00:06:44.500 | Of what the actual systems are doing

00:06:47.500 | That's what the media often reports

00:06:49.500 | That's what some of our speakers will explore in a rigorous way

00:06:56.500 | It's still an important topic to explore

00:06:59.500 | Ray Kurzweil on Wednesday will explore this topic

00:07:03.500 | Next week, talking about AI safety and autonomous weapon systems

00:07:07.500 | Will explore this topic

00:07:09.500 | The future impact 10, 20 years out

00:07:11.500 | How do we design systems today that would lead to safe systems tomorrow?

00:07:15.500 | Still very important

00:07:17.500 | But the reality is a lot of us need to put a lot more emphasis

00:07:21.500 | On the left, on the for loops, on creating these systems

00:07:25.500 | At the same time, the second goal of what we're trying to do here

00:07:29.500 | Is not emphasize the silliness, the simplicity

00:07:33.500 | The naive basic nature of this for loop

00:07:37.500 | In the same way as was the process in creating nuclear weapons

00:07:43.500 | Before, during World War II

00:07:47.500 | The idea that as an engineer, as a scientist

00:07:50.500 | That I'm just a scientist is also a flawed way of thinking

00:07:55.500 | We have to consider the big picture impact

00:07:58.500 | The near term, negative consequences that are preventable

00:08:02.500 | The low hanging fruit that can be prevented through that very engineering process

00:08:07.500 | We have to do both

00:08:09.500 | And in this engineering approach

00:08:13.500 | We always have to be cautious

00:08:16.500 | That just because we don't understand

00:08:20.500 | Just because we, our intuition

00:08:24.500 | Our best understanding of the capabilities of modern systems

00:08:28.500 | That learn, that act in this world

00:08:30.500 | Seem limited, seem far from human level intelligence

00:08:33.500 | Our ability to learn and represent common sense reasoning

00:08:37.500 | Seems limited

00:08:39.500 | The exponential, potentially exponential

00:08:42.500 | Could be argued and he will

00:08:44.500 | Growth of technology, of these ideas

00:08:48.500 | Means that just around the corner is a singularity

00:08:52.500 | Is a breakthrough idea that will change everything

00:08:56.500 | We have to be cautious of that

00:08:58.500 | Moreover, we have to be cautious of the fact that

00:09:04.500 | Every decade over the past century

00:09:07.500 | Our adoption of new technologies has gotten faster and faster

00:09:11.500 | The rate at which a new technology

00:09:14.500 | From its birth to its wide mass adoption

00:09:18.500 | Has shortened and shortened and shortened

00:09:21.500 | That means that new idea

00:09:25.500 | The moment it drops into the world

00:09:28.500 | Can have widespread effects overnight

00:09:32.500 | So as, and I think the engineering approach

00:09:36.500 | Is fundamentally cynical on artificial general intelligence

00:09:39.500 | Because every aspect of it is so difficult

00:09:42.500 | We have to always remember that overnight everything can change

00:09:46.500 | Through this question of beginning to approach

00:09:50.500 | From a deep learning perspective

00:09:52.500 | From deep reinforcement learning

00:09:53.500 | From brain simulation, computational cognitive science

00:09:57.500 | From computational neuroscience

00:10:00.500 | From cognitive architecture, from robotics

00:10:03.500 | From legal perspectives

00:10:07.500 | And autonomous weapon systems

00:10:09.500 | As we begin to approach these questions

00:10:12.500 | We need to start to build intuition

00:10:14.500 | How far away are we from creating intelligent systems

00:10:18.500 | The singularity here is that spark

00:10:21.500 | That moment when we're truly surprised

00:10:25.500 | By the intelligence of the systems we create

00:10:29.500 | I'd like to visualize it

00:10:32.500 | By the certain analogy that we're in this dark room

00:10:37.500 | Looking for a light switch

00:10:39.500 | With no knowledge of where the light switch is

00:10:43.500 | There's going to be people that say

00:10:45.500 | Well, it's a small, the rooms are all small

00:10:47.500 | We're right there

00:10:48.500 | It's anywhere we'll be able to find it in any time

00:10:51.500 | The reality is we know very little

00:10:53.500 | So we have to stumble around

00:10:54.500 | Feel our way around to build the intuition

00:10:57.500 | Of how far away we really are

00:11:03.500 | And many will, speakers here will talk about

00:11:07.500 | How we define intelligence

00:11:08.500 | How we can begin to see intelligence

00:11:11.500 | What are the fundamental impacts

00:11:13.500 | Of creating intelligent systems

00:11:15.500 | I'd like to sort of see the positive reason

00:11:21.500 | For this little class

00:11:23.500 | And for these efforts that have fascinated

00:11:26.500 | People throughout the century

00:11:27.500 | Of trying to create intelligent systems

00:11:30.500 | Is that there's something about human beings

00:11:33.500 | That craves to explore

00:11:39.500 | To uncover the mysteries of the universe

00:11:42.500 | Fundamental in itself a desire

00:11:44.500 | To uncover the mysteries of the universe

00:11:46.500 | Not for a purpose

00:11:48.500 | And there's often an underlying purpose

00:11:50.500 | Of money, of greed, of power

00:11:55.500 | Craving for power and so on

00:11:57.500 | But there's seems to be an underlying desire

00:11:59.500 | To explore

00:12:01.500 | Nice little book, An Exploration

00:12:03.500 | A very short introduction by Stuart Weaver

00:12:05.500 | He says for all the different forms it takes

00:12:09.500 | In different historical periods

00:12:11.500 | For all the worthy and unworthy motives

00:12:14.500 | That lie behind it

00:12:16.500 | Exploration

00:12:19.500 | Travel for the sake of discovery and adventure

00:12:21.500 | Is a human compulsion

00:12:24.500 | A human obsession even

00:12:26.500 | It is defining element

00:12:28.500 | Of a distinctly human identity

00:12:30.500 | And it will never rest at any frontier

00:12:33.500 | Whether terrestrial or extraterrestrial

00:12:37.500 | From 325 BCE

00:12:40.500 | With a long 7500 mile journey

00:12:45.500 | On the ocean

00:12:47.500 | To explore the Arctic

00:12:49.500 | To Christopher Columbus and his

00:12:53.500 | Flawed, harshly criticized

00:12:55.500 | In modern scholarship

00:12:57.500 | Trip that ultimately paved the way

00:12:59.500 | Didn't discover

00:13:01.500 | Paved the way to colonization of the Americas

00:13:08.500 | To the Darwin trip

00:13:10.500 | The voyage of the Beagle

00:13:13.500 | Whilst this planet has gone cycling on

00:13:15.500 | According to the fixed law of gravity

00:13:17.500 | From so simple a beginning

00:13:19.500 | Endless forms most beautiful

00:13:21.500 | And most wonderful have been

00:13:23.500 | And are being evolved

00:13:26.500 | To the first venture into space

00:13:33.500 | By Yuri Gagarin

00:13:35.500 | First human in space in 1961

00:13:40.500 | What he said over the radio

00:13:42.500 | Is the earth is blue

00:13:43.500 | It is amazing

00:13:46.500 | These are the words that I think drive

00:13:49.500 | Our exploration in the sciences

00:13:52.500 | In the engineering

00:13:54.500 | And today in AI

00:13:56.500 | From the first walk on the moon

00:14:00.500 | And now the desire to

00:14:03.500 | Colonize Mars and beyond

00:14:11.500 | That's where I see this desire

00:14:13.500 | To create intelligent systems

00:14:16.500 | Talking about the positive or negative

00:14:18.500 | Impact of AI on society

00:14:20.500 | Talking about the business case

00:14:21.500 | Of the jobs lost, jobs gained, jobs created

00:14:25.500 | Diseases cured

00:14:28.500 | The autonomous vehicles

00:14:30.500 | The ethical questions

00:14:31.500 | The safety of autonomous weapons

00:14:33.500 | Of the misuse of AI

00:14:35.500 | In the financial markets

00:14:37.500 | Underneath it all

00:14:38.500 | And there are people

00:14:39.500 | Many people have spoken about this

00:14:41.500 | What drives myself and many in the community

00:14:44.500 | Is the desire to explore

00:14:46.500 | To uncover the mystery of the universe

00:14:48.500 | And I hope that you join me

00:14:49.500 | In that very effort

00:14:50.500 | With the speakers that come here

00:14:51.500 | In the next two weeks and beyond

00:14:57.500 | The website for the course is

00:14:59.500 | AGI.MIT.EDU

00:15:02.500 | I am a part of an amazing team

00:15:06.500 | Many of whom you know

00:15:09.500 | AGI@MIT.EDU is the email

00:15:12.500 | We're on slack, deep-mit.slack

00:15:18.500 | For registered MIT students

00:15:21.500 | You create an account on the website

00:15:23.500 | And submit five new links

00:15:26.500 | And vote on ten to vote AI

00:15:28.500 | Which is an aggregator of information

00:15:31.500 | And material we've put together

00:15:32.500 | For the topic of AGI

00:15:34.500 | And submit a entry

00:15:38.500 | To one of the competitions

00:15:39.500 | One of the three competitions

00:15:41.500 | And projects that we have in this course

00:15:45.500 | And the projects are

00:15:46.500 | Dream Vision

00:15:47.500 | I'll go over them in a little bit

00:15:48.500 | Dream Vision, Angel, Ethical Car

00:15:52.500 | And the aggregator of material, Vote AI

00:15:55.500 | We have guest speakers

00:15:56.500 | Incredible guest speakers

00:15:57.500 | I will go over them today

00:15:59.500 | And as before

00:16:03.500 | With the deep learning

00:16:04.500 | For self-driving cars course

00:16:05.500 | We have shirts

00:16:07.500 | And they're free for in-person

00:16:10.500 | For people that attend in person

00:16:11.500 | For the last lecture most likely

00:16:13.500 | Or you can order them online

00:16:17.500 | Okay, Dream Vision

00:16:19.500 | We take the Google Deep Dream idea

00:16:22.500 | We explore the idea of creativity

00:16:26.500 | Where Einstein's view of intelligence

00:16:28.500 | The mark of intelligence is creativity

00:16:32.500 | This idea is something we explore

00:16:35.500 | By using neural networks

00:16:36.500 | And interesting ways to visualize

00:16:40.500 | What the network see

00:16:42.500 | And in so doing

00:16:43.500 | Create beautiful visualizations

00:16:45.500 | In time through video

00:16:47.500 | So taking the ideas of deep dream

00:16:50.500 | And combining them together

00:16:52.500 | With multiple video streams

00:16:54.500 | To mix dream and reality

00:16:58.500 | And the competition

00:17:00.500 | Is through Mechanical Turk

00:17:02.500 | We set up a competition of

00:17:05.500 | Who produces the most beautiful visualization

00:17:09.500 | We provide code

00:17:10.500 | To generate this visualization

00:17:12.500 | And ideas of how you can make it

00:17:14.500 | More and more beautiful

00:17:16.500 | And how to submit it

00:17:18.500 | To the competition

00:17:21.500 | ANGEL

00:17:23.500 | The artificial neural generator

00:17:24.500 | Of emotion and language

00:17:27.500 | Is a different twist on the Turing test

00:17:31.500 | Where we don't use words

00:17:34.500 | We only use emotions to speak

00:17:37.500 | Expression of those emotions

00:17:39.500 | And we create

00:17:41.500 | We use an age

00:17:43.500 | A face

00:17:45.500 | Customizable

00:17:47.500 | With 26 muscles

00:17:48.500 | All of which can be controlled

00:17:50.500 | With an LSTM

00:17:52.500 | We use a neural network

00:17:53.500 | To train the generation of emotion

00:17:58.500 | And the competition

00:18:00.500 | In you submitting the code

00:18:03.500 | To the competition

00:18:05.500 | Is you get 10 seconds

00:18:08.500 | To impress

00:18:10.500 | With these expressions of emotion

00:18:12.500 | The viewer

00:18:14.500 | It's A/B testing

00:18:15.500 | Your goal is to impress the viewer

00:18:18.500 | Enough to where they choose

00:18:20.500 | Your agent versus another agent

00:18:23.500 | And those that are most loved

00:18:25.500 | The agents most loved

00:18:27.500 | Will be the ones that are declared winners

00:18:31.500 | In a twist

00:18:32.500 | We will add human beings into this mix

00:18:36.500 | So we've created a system

00:18:38.500 | That maps our human faces

00:18:40.500 | Myself and the TAs

00:18:43.500 | To where we ourselves

00:18:44.500 | Enter in the competition

00:18:47.500 | And try to convince you

00:18:49.500 | To keep us as your friend

00:18:51.500 | That's the Turing test

00:18:57.500 | ETHICAL CAR

00:18:58.500 | Building in the ideas

00:19:00.500 | Of the trolley problem

00:19:01.500 | And the moral machine

00:19:02.500 | Done here in the media lab

00:19:04.500 | The incredible interesting work

00:19:06.500 | We take a machine learning approach to it

00:19:08.500 | And take what we've developed

00:19:10.500 | The deep reinforcement learning competition

00:19:12.500 | For 6S094

00:19:15.500 | The deep traffic

00:19:17.500 | And we add pedestrians into it

00:19:20.500 | Stochastic

00:19:24.500 | Irrational

00:19:25.500 | Unpredictable pedestrians

00:19:28.500 | And we add human life to the loss function

00:19:31.500 | Where there's a trade-off

00:19:33.500 | Between getting from point A to point B

00:19:36.500 | So in deep traffic

00:19:37.500 | The deep reinforcement learning competition

00:19:39.500 | The goal was to go as fast as possible

00:19:41.500 | Here it's up to you

00:19:43.500 | To decide

00:19:45.500 | What your agent's goal is

00:19:47.500 | There's a parade of front trade-off

00:19:50.500 | Between getting from point A to point B

00:19:52.500 | As fast as possible

00:19:54.500 | And

00:19:58.500 | Hurting pedestrians

00:20:08.500 | This is not a ethical question

00:20:11.500 | It's an engineering question

00:20:17.500 | And it's a serious one

00:20:19.500 | Because fundamentally in creating

00:20:22.500 | Autonomous vehicles that function in this world

00:20:25.500 | We want them to get

00:20:27.500 | From point A to point B as quickly as possible

00:20:31.500 | The United States government

00:20:33.500 | Insurance companies

00:20:35.500 | Put a price tag on human life

00:20:38.500 | We put that power in your hands

00:20:40.500 | In designing these agents

00:20:42.500 | To ask the question of

00:20:44.500 | How can we create machine learning systems

00:20:48.500 | Where the objective function

00:20:49.500 | The loss function

00:20:50.500 | Has human life as part of it

00:20:53.500 | And Vote.ai is an aggregator

00:20:58.500 | Of different links

00:21:01.500 | Different articles, papers, videos

00:21:04.500 | On the topic of artificial general intelligence

00:21:06.500 | Where people vote on

00:21:08.500 | Vote quality articles up and down

00:21:12.500 | And choose on the sentiment

00:21:15.500 | Of positive and negative

00:21:17.500 | We'd like to explore the different ways

00:21:19.500 | The different arguments for and against

00:21:21.500 | Artificial general intelligence

00:21:26.500 | There is an incredible list of speakers

00:21:29.500 | The best in their disciplines

00:21:32.500 | From Josh Tenenbaum here at MIT

00:21:35.500 | To Ray Kurzweil at Google

00:21:37.500 | To Lisa Feldman Barrett

00:21:39.500 | And Nate Dzerbinski from Northeastern University

00:21:43.500 | Andrej Karpathy

00:21:45.500 | Stephen Wolfram

00:21:47.500 | Richard Moyes

00:21:48.500 | Mark Reiber

00:21:49.500 | Ilyas Eskiver

00:21:52.500 | And myself

00:21:54.500 | Josh Tenenbaum, tomorrow

00:21:56.500 | I'd like to go through each of these speakers

00:21:59.500 | And talk about the perspectives they bring

00:22:03.500 | To try to see the approach

00:22:06.500 | The ideas they bring to the table

00:22:08.500 | They're not, in most cases, interested

00:22:11.500 | In the discussion of the future impact on society

00:22:17.500 | Without grounding it into their expertise

00:22:20.500 | Into their actual engineering

00:22:22.500 | Into creating these intelligence systems

00:22:24.500 | So Josh is a computational cognitive science

00:22:28.500 | Expert professor faculty here at MIT

00:22:31.500 | He will talk about

00:22:33.500 | How we can create common sense understanding

00:22:35.500 | Systems that

00:22:37.500 | See a world of physical objects

00:22:39.500 | And their interactions

00:22:41.500 | And our own possibilities to act and interact with others

00:22:44.500 | The intuitive physics

00:22:45.500 | How do we build into systems

00:22:47.500 | The intuitive physics of the world

00:22:49.500 | More than just the deep learning memorization engines

00:22:53.500 | That take patterns

00:22:55.500 | And learn through supervised way

00:22:57.500 | To map those patterns to classification

00:23:00.500 | Actually begin to understand the intuitive

00:23:03.500 | The common sense physics of the world

00:23:06.500 | And learn rapid model-based learning

00:23:09.500 | Learn from nothing

00:23:10.500 | Learn from very little

00:23:11.500 | Just like we do as children

00:23:13.500 | Just like we do as human beings successfully

00:23:15.500 | Often only need one example to learn a concept

00:23:18.500 | How do we create systems

00:23:20.500 | That learn from very few

00:23:23.500 | Sometimes a single example

00:23:25.500 | And integrate ideas from various disciplines

00:23:28.500 | Of course from neural networks

00:23:30.500 | But also probabilistic generative models

00:23:33.500 | And simple processing architectures

00:23:36.500 | It's gonna be incredible

00:23:38.500 | Of course from a different area of the world

00:23:42.500 | Another incredible thinker

00:23:45.500 | Intellectual speaker is Ray Kurzweil

00:23:48.500 | He'll be here on Wednesday at 1pm

00:23:51.500 | And he will do a whirlwind discussion

00:23:56.500 | Of where we stand with intelligence

00:23:59.500 | Creating intelligence systems

00:24:00.500 | How we see natural intelligence

00:24:02.500 | Our own human intelligence

00:24:03.500 | How we define it

00:24:05.500 | How we understand it

00:24:06.500 | And how that transfers to the increasing exponential growth

00:24:11.500 | Of development of artificial general intelligence

00:24:16.500 | Something I'm myself very excited about

00:24:20.500 | Is Lisa Feldman Barrett coming here on Thursday

00:24:25.500 | She's written a book

00:24:27.500 | I believe "How Emotions Are Made"

00:24:30.500 | She argues that emotions are created

00:24:34.500 | That there is a distinction

00:24:36.500 | There is a detachment between what we feel in our bodies

00:24:40.500 | The physical state of our bodies

00:24:42.500 | And the expression of emotion

00:24:44.500 | From body to the contextually grounded

00:24:48.500 | To the face expressing that emotion

00:24:50.500 | Which means

00:24:52.500 | Now why is there a person who is a psychology person

00:24:56.500 | In a fundamental engineering computer science topic like AGI

00:25:00.500 | Because if emotions are created in the way she argues

00:25:04.500 | And she'll systematically break it down

00:25:06.500 | That means we're learning societal

00:25:10.500 | As human beings we're learning societal norms

00:25:12.500 | Of how to express emotion

00:25:14.500 | The idea of emotional intelligence is learned

00:25:16.500 | Which means we can have machines learn this idea

00:25:20.500 | It's a machine learn

00:25:21.500 | Just like it's a human learning problem

00:25:23.500 | It's a machine learning problem

00:25:26.500 | In a little bit of a twist

00:25:28.500 | She asked that instead of giving a talk

00:25:30.500 | I have a conversation with her

00:25:33.500 | So it's going to be a little bit challenging and fun

00:25:36.500 | And she's great looking forward to it

00:25:39.500 | And we'll explore different ways

00:25:42.500 | That we can get emotion expressed

00:25:46.500 | Through video, through audio, through the project

00:25:50.500 | The Angel project that I mentioned

00:25:52.500 | So there's been work in reenacting intelligence

00:25:55.500 | So well reenacting mapping face to face

00:25:59.500 | Mapping different emotions on video that was previously recorded

00:26:03.500 | So if you can imagine

00:26:05.500 | That means we can take emotions that we've created

00:26:09.500 | The kind of emotion creation we've been discussing

00:26:12.500 | And remap it on previous video

00:26:16.500 | That's one way to see intelligence

00:26:18.500 | Is taking raw human data that we already have

00:26:21.500 | And mapping new computer generated

00:26:24.500 | The underlying fundamentals of human

00:26:28.500 | But the surface appearance

00:26:30.500 | The representation of emotion visual or auditory

00:26:35.500 | Is generated by computer

00:26:39.500 | It could be in the embodied form like with Sophia

00:26:43.500 | Sophia, your robots were human

00:26:48.500 | I think we will be similar in a lot of ways

00:26:52.500 | But different in a few others

00:26:54.500 | It will take a long time for robots

00:26:56.500 | To develop a complex emotion

00:26:58.500 | And possibly robots can be able to

00:27:00.500 | Develop the more problematic emotions

00:27:02.500 | Like rage, jealousy, hatred and so on

00:27:06.500 | It might be possible to make them more effective

00:27:08.500 | In humans

00:27:10.500 | So I think it will be a good partnership

00:27:12.500 | Where one of them completes the other

00:27:14.500 | Very important to note

00:27:19.500 | For those captivated by Sophia in the press

00:27:22.500 | Or have seen these videos

00:27:24.500 | Sophia is an art exhibit

00:27:27.500 | She's not a strong natural language processing system

00:27:32.500 | This is not an AGI system

00:27:34.500 | But it's a beautiful visualization

00:27:38.500 | Of embodying

00:27:39.500 | It's a beautiful visualization

00:27:42.500 | Of how easy it is to trick us human beings

00:27:44.500 | That there's intelligence underlying something

00:27:47.500 | That the emotional expression

00:27:50.500 | The physical embodiment

00:27:51.500 | And the emotional expression

00:27:53.500 | That has some degree of humor

00:27:57.500 | That has some degree of wit and intelligence

00:28:00.500 | Is enough to captivate us

00:28:02.500 | So that's an argument for

00:28:04.500 | Not creating intelligence from scratch

00:28:06.500 | But having machines at the very surface

00:28:09.500 | The display of that emotion

00:28:11.500 | The generation, the mapping of the visual

00:28:14.500 | And the auditory elements

00:28:16.500 | Where underneath it is really trivial technology

00:28:20.500 | That's fundamentally relying on humans

00:28:22.500 | Like in the Sophia's case

00:28:24.500 | And in the simplest form

00:28:27.500 | We remove all elements of

00:28:30.500 | How should I say

00:28:32.500 | Attractive appearance from an agent

00:28:35.500 | We really keep it to the simplest muscles

00:28:38.500 | Aspect characteristics of the face

00:28:40.500 | And see with 26 muscles

00:28:42.500 | Controlled by a neural network

00:28:43.500 | Through time

00:28:44.500 | So a current neural network LSTM

00:28:46.500 | How can we explore the generation of emotion

00:28:50.500 | Can we get this thing

00:28:51.500 | And this is an open question for us too

00:28:53.500 | We just created this system

00:28:54.500 | We don't know if we can

00:28:55.500 | Can we get it to make us feel something

00:28:59.500 | Make us feel something by

00:29:01.500 | Watching it express its feelings

00:29:06.500 | Can it become human before our eyes

00:29:08.500 | Can it learn to

00:29:09.500 | By competing against other agents

00:29:12.500 | A, B testing on Turk

00:29:15.500 | On Mechanical Turk

00:29:17.500 | Can the winners be very convincing

00:29:19.500 | To make us feel entertained

00:29:24.500 | Pity

00:29:26.500 | Love

00:29:27.500 | Maybe some of you will fall in love

00:29:28.500 | With Angel here

00:29:32.500 | Nate Derbenski on Friday

00:29:34.500 | Will talk about

00:29:35.500 | Cognitive modeling architectures

00:29:37.500 | So he will speak about

00:29:38.500 | The cognitive modeling aspect

00:29:40.500 | Can we have a

00:29:42.500 | Can we model cognition

00:29:44.500 | In some kind of systematic way

00:29:46.500 | To try to build intuition

00:29:48.500 | Of how complicated cognition is

00:29:53.500 | Andrej Karpathy

00:29:56.500 | Famous for being the state of the art human

00:29:59.500 | On the ImageNet challenge

00:30:01.500 | The representative

00:30:02.500 | The 95% accuracy performance

00:30:06.500 | Among other things

00:30:07.500 | He's also famous for

00:30:08.500 | Is now at Tesla

00:30:10.500 | He will talk about

00:30:12.500 | The role

00:30:14.500 | The limitations

00:30:15.500 | The possibilities of deep learning

00:30:19.500 | We'll talk

00:30:21.500 | As I have spoken about

00:30:24.500 | In the past few weeks

00:30:25.500 | And throughout

00:30:26.500 | About our misunderstanding

00:30:29.500 | Or our flawed intuition

00:30:31.500 | About what are the difficult

00:30:32.500 | And what are the easy problems

00:30:34.500 | In deep learning

00:30:36.500 | And the power

00:30:37.500 | Of the representational learning

00:30:39.500 | The ability of neural networks

00:30:41.500 | To form deeper and deeper representations

00:30:43.500 | Of the underlying raw data

00:30:45.500 | That ultimately forms

00:30:48.500 | That takes complex information

00:30:50.500 | That's hard to make sense of

00:30:52.500 | And convert it into useful

00:30:54.500 | Actionable knowledge

00:30:57.500 | That is

00:30:59.500 | From a certain lens

00:31:01.500 | In a certain

00:31:03.500 | In a certain lens

00:31:04.500 | In a certain problem space

00:31:06.500 | Can be clearly defined

00:31:08.500 | As understanding

00:31:10.500 | Of the complex information

00:31:11.500 | Understanding is ultimately

00:31:13.500 | Taking complex information

00:31:15.500 | And reducing it to

00:31:16.500 | Its simple essential elements

00:31:19.500 | Representational learning

00:31:21.500 | In the trivial case here

00:31:23.500 | In drawing

00:31:24.500 | Having to draw a straight line

00:31:26.500 | To separate the blue and the red curves

00:31:29.500 | That's impossible to do

00:31:31.500 | In a natural input space on the left

00:31:34.500 | What the act of learning is

00:31:36.500 | For deep neural networks

00:31:37.500 | In this formulation

00:31:39.500 | Is to construct a topology

00:31:40.500 | Under which there exists a straight line

00:31:43.500 | To accurately classify blue versus red

00:31:46.500 | That's the problem

00:31:48.500 | And for a simple blue and red line

00:31:50.500 | It seems trivial here

00:31:51.500 | But this works in a general case

00:31:53.500 | For arbitrary input spaces

00:31:55.500 | For arbitrary nonlinear

00:31:56.500 | Highly dimensional input spaces

00:31:59.500 | And the ability to automatically learn features

00:32:02.500 | To learn hierarchical representations

00:32:07.500 | Of the raw sensory data

00:32:09.500 | Which means that you could do

00:32:10.500 | A lot more with data

00:32:11.500 | Which means you can expand

00:32:13.500 | Further and further and further

00:32:15.500 | To create intelligent systems

00:32:17.500 | That operate successfully

00:32:19.500 | With real world data

00:32:20.500 | That's what representational learning means

00:32:22.500 | That deep learning allows

00:32:24.500 | Because the arbitrary number of features

00:32:26.500 | That can be automatically determined

00:32:28.500 | You can learn a lot of things

00:32:30.500 | About a pretty complex world

00:32:32.500 | Unfortunately

00:32:34.500 | There needs to be a lot of supervised data

00:32:36.500 | There still needs to be a lot of human input

00:32:41.500 | Andre and others, Josh

00:32:43.500 | Will talk about the difference

00:32:45.500 | Between our human brain

00:32:47.500 | Our biological neural network

00:32:49.500 | And the artificial neural network

00:32:52.500 | The full human brain

00:32:53.500 | With hundred billion neurons

00:32:55.500 | One thousand trillion synapses

00:32:58.500 | And the biggest neural networks out there

00:33:01.500 | The artificial neural networks

00:33:03.500 | Having much smaller

00:33:05.500 | 60 million synapses for ResNet-152

00:33:09.500 | The biggest difference

00:33:11.500 | The parameters of human brain

00:33:13.500 | Being several orders of magnitude

00:33:15.500 | More synapses

00:33:17.500 | The topology being much more complex

00:33:20.500 | Chaotic

00:33:21.500 | The asynchronous nature of the human brain

00:33:24.500 | And the learning algorithm

00:33:25.500 | Of artificial neural networks

00:33:27.500 | Is trivial and constrained

00:33:30.500 | With back propagation

00:33:31.500 | Is essentially an optimization function

00:33:33.500 | Over a clearly defined loss function

00:33:37.500 | From the output to the input

00:33:40.500 | Using back propagation to teach

00:33:42.500 | To adjust the weights on that network

00:33:44.500 | The learning algorithm

00:33:46.500 | For our human brain

00:33:48.500 | Is mostly unknown

00:33:50.500 | But it's certainly much more complicated

00:33:52.500 | Than back propagation

00:33:54.500 | The power consumption

00:33:57.500 | The human brain is a lot more efficient

00:33:59.500 | Than artificial neural networks

00:34:01.500 | And there's a very kind of artificial

00:34:04.500 | Trivial supervised learning process

00:34:09.500 | For training artificial neural networks

00:34:11.500 | You have to have a training stage

00:34:13.500 | And you have to have an evaluation stage

00:34:15.500 | And once the network is trained

00:34:17.500 | There's no clear way to continue training it

00:34:19.500 | Or there's a lot of ways

00:34:21.500 | But they're inefficient

00:34:22.500 | It's not designed to do online learning

00:34:25.500 | Naturally

00:34:27.500 | To always be learning

00:34:29.500 | Is designed to be to learn

00:34:31.500 | And then be applied

00:34:33.500 | Obviously our human brains are always learning

00:34:36.500 | But the beautiful fascinating thing is

00:34:39.500 | That they're both distributed

00:34:41.500 | Computation systems on a large scale

00:34:43.500 | So it's not a

00:34:46.500 | There's, it doesn't ultimately boil down

00:34:48.500 | To a single compute unit

00:34:51.500 | The computation is distributed

00:34:53.500 | The back propagation learning process

00:34:55.500 | Is distributed

00:34:56.500 | Can be paralyzed on a GPU

00:34:58.500 | Massively paralyzed

00:35:00.500 | The underlying computational unit of a neuron

00:35:03.500 | Is trivial

00:35:04.500 | But can be stacked together

00:35:06.500 | To form forward neural networks

00:35:07.500 | Recurrent neural networks

00:35:09.500 | To represent both spatial information with images

00:35:13.500 | And temporal information

00:35:16.500 | With audio speech

00:35:19.500 | Text

00:35:20.500 | Sequences of images and video

00:35:23.500 | And so on

00:35:25.500 | Mapping from one to one

00:35:26.500 | One to many

00:35:27.500 | Many to one

00:35:28.500 | So mapping any kind of structure

00:35:31.500 | Vector and time data

00:35:33.500 | As an input

00:35:34.500 | To any kind of classification regression

00:35:37.500 | Sequences

00:35:38.500 | Captioning

00:35:39.500 | Video

00:35:40.500 | Audio as output

00:35:41.500 | Learning in the general sense

00:35:44.500 | But in a domain

00:35:46.500 | That's precisely defined

00:35:48.500 | For the supervised training process

00:35:52.500 | We can think of

00:35:55.500 | In deep learning case

00:35:57.500 | You can think of the supervised methods

00:35:59.500 | Where humans have to annotate the data

00:36:02.500 | As memorization of the data

00:36:04.500 | We can think of the exciting new

00:36:06.500 | And growing field of

00:36:07.500 | Semi-supervised learning

00:36:09.500 | When most of the data

00:36:10.500 | Through generative adversarial networks

00:36:13.500 | Or through significant data augmentation

00:36:16.500 | Clever data augmentation

00:36:17.500 | Most of it is done automatically

00:36:19.500 | The annotation process

00:36:20.500 | Or through simulation

00:36:22.500 | And then reinforcement learning

00:36:23.500 | Where most of the

00:36:25.500 | Most of the labels are extremely sparse

00:36:28.500 | And come rarely

00:36:29.500 | And so the system has to figure out

00:36:31.500 | How to operate in the world

00:36:32.500 | With very little human input

00:36:33.500 | Very little human data

00:36:36.500 | We can think of that as reasoning

00:36:40.500 | Because you take very little information

00:36:41.500 | From our teachers

00:36:42.500 | The humans

00:36:43.500 | And transfer it across

00:36:44.500 | Generalize it across

00:36:46.500 | To reason about the world

00:36:48.500 | And finally unsupervised learning

00:36:50.500 | The excitement of the community

00:36:52.500 | The promise, the hope

00:36:54.500 | You could think of that as understanding

00:36:56.500 | Because ultimately it's taking data

00:36:58.500 | With very little or no human input

00:37:00.500 | And forming representations of that data

00:37:03.500 | Is how we think of understanding

00:37:04.500 | Requiring making sense of the world

00:37:08.500 | Without strict input

00:37:10.500 | Of how to make sense of the world

00:37:13.500 | The kind of process

00:37:14.500 | Of discovering information

00:37:17.500 | Maybe discovering new ideas

00:37:19.500 | New ways to simplify the world

00:37:21.500 | To represent the world

00:37:22.500 | That you can do new things with it

00:37:24.500 | The new is the key element there

00:37:26.500 | Understanding

00:37:28.500 | And Andre and Ilya

00:37:31.500 | And others will talk about

00:37:33.500 | Certainly the past

00:37:34.500 | But the future of deep learning

00:37:36.500 | Where is it going to go?

00:37:38.500 | Is it overhyped?

00:37:40.500 | Underhyped?

00:37:41.500 | What is the future?

00:37:42.500 | Will the compute of CPU, GPU, ASICs

00:37:46.500 | Continue?

00:37:47.500 | Will the breakthroughs

00:37:48.500 | The Moore's law

00:37:49.500 | And its various forms

00:37:50.500 | Of massive parallelization continue?

00:37:52.500 | And the large data sets

00:37:54.500 | With tens of millions of images

00:37:56.500 | Grow to billions and trillions

00:37:59.500 | Will the algorithms improve?

00:38:01.500 | Is there a groundbreaking idea

00:38:03.500 | That's still coming?

00:38:05.500 | With Jeff Hiddens' capsule networks

00:38:07.500 | Is there fundamental architectural changes

00:38:10.500 | To neural networks

00:38:11.500 | That we can come up with

00:38:13.500 | That will change everything

00:38:14.500 | That will ease the learning process

00:38:16.500 | That will make the learning process

00:38:17.500 | More efficient

00:38:19.500 | Or we'll be able to represent

00:38:21.500 | Higher and higher orders of information

00:38:24.500 | Such that you can

00:38:26.500 | Transfer knowledge between domains

00:38:29.500 | And the software architectures

00:38:31.500 | That support from TensorFlow to PyTorch

00:38:34.500 | I would say the last year

00:38:36.500 | And this year will be the year

00:38:37.500 | Of deep learning frameworks

00:38:39.500 | Those will certainly keep coming

00:38:40.500 | In their various forms

00:38:42.500 | And the financial backing

00:38:44.500 | Is growing and growing

00:38:47.500 | The open challenges for deep learning

00:38:50.500 | Really a lot of this course

00:38:52.500 | Is kind of connected to deep learning

00:38:54.500 | Because that's where a lot

00:38:56.500 | Of the recent breakthroughs

00:38:58.500 | That inspire us

00:39:00.500 | To think about intelligence systems

00:39:02.500 | Come from

00:39:03.500 | But the challenges are many

00:39:05.500 | The need, the ability to transfer

00:39:07.500 | Between different domains

00:39:09.500 | As in reinforcement learning and robotics

00:39:11.500 | The need for huge data

00:39:13.500 | And inefficient learning

00:39:15.500 | We still need supervised data

00:39:18.500 | An ability to learn

00:39:20.500 | In an unsupervised way

00:39:22.500 | Is a huge problem

00:39:23.500 | And not fully automated learning

00:39:26.500 | There's still a degree

00:39:27.500 | A significant degree of hyperparameter

00:39:29.500 | To necessary

00:39:30.500 | With the reward functions

00:39:32.500 | The loss functions

00:39:33.500 | Are ultimately defined by humans

00:39:36.500 | And therefore are deeply flawed

00:39:38.500 | When we release those systems

00:39:40.500 | Into the real world

00:39:41.500 | Where there is no ground truth

00:39:43.500 | For the testing set

00:39:45.500 | And the goal isn't achieving

00:39:47.500 | A high classification

00:39:49.500 | On a trivial image classification

00:39:52.500 | Localization detection problem

00:39:54.500 | But rather to have an autonomous vehicle

00:39:56.500 | That doesn't kill pedestrians

00:39:58.500 | Or an industrial robot

00:40:02.500 | That operates in jointly

00:40:04.500 | With other human beings

00:40:05.500 | And all the edge cases that come up

00:40:08.500 | How does deep learning methods

00:40:10.500 | How do machine learning methods

00:40:11.500 | Generalize over the edge cases

00:40:12.500 | The weird stuff that happens

00:40:13.500 | In the real world

00:40:14.500 | Those are all the problems there

00:40:16.500 | Stephen Wolfram will be here on Monday

00:40:19.500 | Evening at 7 p.m.

00:40:22.500 | He's done a lot of amazing things

00:40:25.500 | I would say

00:40:26.500 | It's very interesting

00:40:27.500 | From his recent interest

00:40:29.500 | In knowledge-based programming

00:40:31.500 | Wolfram Alpha

00:40:32.500 | I think is the fuel

00:40:34.500 | For most middle school

00:40:36.500 | And high school students now

00:40:38.500 | For the first time taking calculus

00:40:41.500 | I probably go to Wolfram Alpha

00:40:43.500 | To answer their own questions

00:40:44.500 | But more seriously

00:40:46.500 | There is a deep connected graph

00:40:49.500 | Of knowledge that's being built there

00:40:51.500 | With Wolfram Alpha

00:40:53.500 | And Wolfram language

00:40:55.500 | That Stephen will explore

00:40:57.500 | In terms of language

00:40:59.500 | An interesting thing

00:41:01.500 | He was part of the team on Arrival

00:41:04.500 | That worked on the language

00:41:06.500 | For those of you who are familiar

00:41:08.500 | The Arrival where alien species spoke with

00:41:11.500 | Us humans

00:41:14.500 | Through a very interesting

00:41:16.500 | Beautiful complicated language

00:41:18.500 | And he was brought in

00:41:19.500 | As a representative human

00:41:21.500 | To interpret that language

00:41:23.500 | Just like in the movie

00:41:25.500 | He's represented that in real life

00:41:27.500 | And he used the skills

00:41:29.500 | Him and his son Christopher

00:41:31.500 | Used to analyze this language

00:41:33.500 | Very interesting

00:41:34.500 | That process is extremely interesting

00:41:35.500 | I hope he talks about it

00:41:37.500 | And his background with Mathematica

00:41:39.500 | And new kind of science

00:41:42.500 | The sort of

00:41:45.500 | Another set of ideas

00:41:48.500 | That have inspired people

00:41:50.500 | In terms of creating

00:41:53.500 | Intelligence systems

00:41:55.500 | Is the idea that

00:41:59.500 | From very simple things

00:42:02.500 | From very simple rules

00:42:04.500 | Extremely complex

00:42:06.500 | Patterns can emerge

00:42:08.500 | His work with cellular automata

00:42:10.500 | Did just that

00:42:11.500 | Taking extremely simple

00:42:13.500 | Mathematical constructs

00:42:16.500 | Here with cellular automata

00:42:18.500 | These are

00:42:19.500 | These are grids

00:42:21.500 | Of computational units

00:42:22.500 | That switch on and off

00:42:24.500 | In some kind of predefined way

00:42:26.500 | And only operate locally

00:42:27.500 | Based on their local neighborhood

00:42:29.500 | And somehow based on

00:42:30.500 | Different kinds of rules

00:42:32.500 | Different patterns emerge

00:42:33.500 | Here's a three-dimensional cellular automata

00:42:35.500 | With a simple rule

00:42:36.500 | Starting with nothing

00:42:37.500 | With a single cell

00:42:38.500 | They grow in really

00:42:40.500 | Interesting complex ways

00:42:41.500 | This emergent complexity

00:42:43.500 | Is inspiring

00:42:45.500 | It's the same kind of

00:42:46.500 | Thing that inspires us

00:42:47.500 | About neural networks

00:42:49.500 | You can take a simple

00:42:50.500 | Computational unit

00:42:51.500 | And when combined together

00:42:52.500 | In arbitrary ways

00:42:53.500 | Can form complex representations

00:42:56.500 | That's also very interesting

00:42:59.500 | You can see knowledge

00:43:00.500 | From a knowledge perspective

00:43:01.500 | You can see knowledge formation

00:43:02.500 | In the same kind of way

00:43:04.500 | Simplicity

00:43:06.500 | At a mass distributed scale

00:43:09.500 | Resulting in complexity

00:43:11.500 | Next Tuesday

00:43:13.500 | Richard Moyes from article 36

00:43:16.500 | Coming all the way from UK

00:43:18.500 | For us

00:43:19.500 | We'll talk about

00:43:20.500 | It works with autonomous weapon systems

00:43:23.500 | Works with also

00:43:25.500 | Nuclear weapons

00:43:26.500 | But primarily autonomous weapon systems

00:43:29.500 | And concern legal policy

00:43:32.500 | And technological aspects

00:43:34.500 | Of banning these weapons

00:43:35.500 | There's been a lot of agreement

00:43:37.500 | About the safety hazards

00:43:39.500 | Of autonomous systems

00:43:41.500 | That make decisions

00:43:42.500 | To kill a human being

00:43:44.500 | Mark Kriber

00:43:46.500 | CEO of Boston Dynamics

00:43:48.500 | Previously

00:43:50.500 | A long time ago

00:43:51.500 | Faculty here at MIT

00:43:53.500 | We'll talk about

00:43:54.500 | We'll bring robots

00:43:55.500 | And talk to us about

00:43:57.500 | His work

00:43:58.500 | Of robots in the real world

00:44:01.500 | Is doing a lot of exciting stuff

00:44:03.500 | With humanoid robotics

00:44:04.500 | And any kind of robots

00:44:05.500 | Operating on legs

00:44:07.500 | It's incredible work

00:44:08.500 | Extremely exciting

00:44:10.500 | And gets to explore the idea

00:44:12.500 | Of how difficult it is

00:44:13.500 | To build these robot systems

00:44:14.500 | That operate in the real world

00:44:17.500 | From both the control aspect

00:44:21.500 | And from the way

00:44:24.500 | The final result is perceived

00:44:26.500 | By our society

00:44:28.500 | It's very interesting

00:44:29.500 | To see what intelligence

00:44:31.500 | In robotics is embodied

00:44:33.500 | And then taking in by us

00:44:35.500 | And what that inspires

00:44:37.500 | Fear, excitement, hope, concern

00:44:41.500 | And all the above

00:44:43.500 | Ilya Siskeev is

00:44:46.500 | Expert in many aspects

00:44:48.500 | Of machine learning

00:44:49.500 | He's the co-founder of OpenAI

00:44:52.500 | I'll talk about their

00:44:55.500 | Different aspects of game playing

00:44:58.500 | That they've recently been exploring

00:45:00.500 | About using

00:45:01.500 | Deep reinforcement learning

00:45:02.500 | To play arcade games

00:45:04.500 | On the DeepMind side

00:45:08.500 | Using deep reinforcement learning

00:45:10.500 | To beat the best in the world

00:45:11.500 | At the game of Go

00:45:13.500 | In 2017

00:45:15.500 | The big fascinating breakthrough

00:45:17.500 | Achieved by that team

00:45:18.500 | With AlphaGo Zero

00:45:19.500 | Training an agent

00:45:20.500 | That through self-play

00:45:21.500 | Playing itself

00:45:22.500 | Not on expert games

00:45:24.500 | So truly from scratch

00:45:26.500 | Learning to beat

00:45:27.500 | The best in the world

00:45:28.500 | Including the previous

00:45:29.500 | Iteration of AlphaGo

00:45:31.500 | We'll explore

00:45:32.500 | What aspects of the stack

00:45:35.500 | Of intelligent robotic systems

00:45:37.500 | Intelligent agents

00:45:38.500 | Can be learned in this way

00:45:40.500 | So deep learning

00:45:41.500 | The memorization

00:45:42.500 | The supervised learning

00:45:44.500 | Memorization approach

00:45:45.500 | It looks at the

00:45:47.500 | Sensor data feature extraction

00:45:49.500 | Representation learning aspect of this

00:45:52.500 | Taking the sensor data

00:45:53.500 | From camera, lidar, audio

00:45:55.500 | Extracting the features

00:45:57.500 | Forming higher order representations

00:46:00.500 | And on those representations

00:46:01.500 | Learning to actually accomplish

00:46:02.500 | Some kind of classification

00:46:04.500 | Regression task

00:46:05.500 | Figuring out

00:46:07.500 | Based on the representation

00:46:08.500 | What is going on

00:46:09.500 | In the raw sensory data

00:46:11.500 | And then combining that data together

00:46:13.500 | To reason about it

00:46:15.500 | And finally in the robotic domains

00:46:18.500 | Taking it all together

00:46:19.500 | As with humanoid robotics

00:46:22.500 | Industrial robotics

00:46:23.500 | Autonomous vehicles

00:46:24.500 | Taking it all together

00:46:25.500 | And actually acting in this world

00:46:27.500 | With the effectors

00:46:28.500 | And the open question is

00:46:31.500 | How much of this AI stack

00:46:32.500 | Can be learned?

00:46:34.500 | That's something for us to discuss

00:46:37.500 | To think about

00:46:38.500 | That Elio will touch on

00:46:40.500 | With deeper enforcement learning

00:46:42.500 | We can certainly learn representations

00:46:44.500 | And perform classifications

00:46:46.500 | State of the art

00:46:47.500 | Better than human

00:46:48.500 | And image classification

00:46:49.500 | And ImageNet

00:46:50.500 | And segmentation tasks

00:46:53.500 | And the excitement of deep learning

00:46:55.500 | Is what's highlighted there in the red box

00:46:57.500 | Can be done end to ends

00:46:58.500 | Raw sensory data

00:46:59.500 | Out to the knowledge

00:47:00.500 | To the output

00:47:01.500 | To the classification

00:47:02.500 | Can we begin to reason?

00:47:04.500 | Is the open question

00:47:05.500 | With the knowledge-based programming

00:47:08.500 | That Stephen Wolfram will talk about

00:47:09.500 | Can we begin to

00:47:11.500 | Take these automatically generated

00:47:13.500 | High order representations

00:47:15.500 | And combine them together

00:47:17.500 | To form knowledge bases

00:47:19.500 | To form aggregate graphs of ideas

00:47:22.500 | That can then be used to reason

00:47:24.500 | And can we then combine them together

00:47:28.500 | To act in the world

00:47:31.500 | For whether in simulation

00:47:32.500 | With arcade games

00:47:34.500 | Or simulation of autonomous vehicles

00:47:36.500 | Robotic systems

00:47:37.500 | Or actually in the physical world

00:47:38.500 | With robots moving about

00:47:40.500 | Can that end to end

00:47:42.500 | From raw sensory data

00:47:43.500 | To action be learned?

00:47:46.500 | That's the open question

00:47:48.500 | For artificial general intelligence

00:47:51.500 | For this class

00:47:52.500 | Can this entire process

00:47:55.500 | Be end to end?

00:47:57.500 | Can we build systems

00:47:59.500 | And how do we do it

00:48:01.500 | That achieve this process end to end

00:48:03.500 | In the same way that humans do?

00:48:05.500 | We're born in this raw sensory environment

00:48:07.500 | Taking in very little information

00:48:10.500 | And learn to operate successfully

00:48:13.500 | In arbitrary constraints

00:48:17.500 | Arbitrary goals

00:48:19.500 | And to do so

00:48:21.500 | We have lectures

00:48:22.500 | We have three projects

00:48:24.500 | And we have guest speakers

00:48:27.500 | From various disciplines

00:48:29.500 | I hope that all these voices

00:48:31.500 | Will be heard

00:48:33.500 | And will feed a conversation

00:48:35.500 | About artificial intelligence

00:48:38.500 | And its positive

00:48:41.500 | And its concerning effects in society

00:48:44.500 | And how do we move forward

00:48:45.500 | From an engineering approach

00:48:47.500 | The topics will be deep learning

00:48:49.500 | Deep reinforcement learning

00:48:51.500 | Cognitive modeling

00:48:52.500 | Computational cognitive science

00:48:54.500 | Emotion creation

00:48:55.500 | Knowledge based programming

00:48:56.500 | AI safety with autonomous weapon systems

00:48:59.500 | And personal robotics

00:49:01.500 | With human centered artificial intelligence

00:49:03.500 | That's for the first two weeks of this class

00:49:05.500 | That's the part

00:49:07.500 | Where if you're actually registered students

00:49:09.500 | That's where you need to submit the project

00:49:11.500 | That's when we all meet here

00:49:13.500 | Every night with the incredible speakers

00:49:15.500 | But this will continue

00:49:17.500 | We already have several speakers scheduled

00:49:19.500 | In the next couple of months

00:49:21.500 | Yet to be announced

00:49:22.500 | But they're incredible

00:49:24.500 | And we have conversations

00:49:26.500 | On video

00:49:27.500 | And we have new projects

00:49:29.500 | I hope this continues throughout 2018

00:49:32.500 | On the topics of AI ethics

00:49:34.500 | And bias

00:49:35.500 | There's a lot of incredible work

00:49:37.500 | In, we have a speaker there coming

00:49:40.500 | On the topic of

00:49:41.500 | How do we create artificial intelligence systems

00:49:43.500 | That do not discriminate

00:49:45.500 | Do not form the kind of biases

00:49:47.500 | That us humans do in this world

00:49:50.500 | That are operating under social norms

00:49:53.500 | But are reasoning beyond the flawed aspects

00:49:57.500 | Of those social norms

00:49:59.500 | With bias

00:50:01.500 | Creativity as with the project

00:50:03.500 | Of Dream Vision and beyond

00:50:05.500 | There's so much exciting work

00:50:07.500 | In using machine learning methods

00:50:10.500 | To create beautiful art and music

00:50:13.500 | Brain simulation, neuroscience

00:50:17.500 | Computational neuroscience

00:50:18.500 | Shockingly, in the first two weeks

00:50:20.500 | We don't have a computational neuroscience speaker

00:50:23.500 | Which is a fascinating perspective

00:50:26.500 | Brain simulation or neuroscience in general

00:50:30.500 | Computational neuroscience

00:50:31.500 | Is a fascinating approach

00:50:33.500 | From the muck of actual brain work

00:50:37.500 | To get the perspective of how our brain works

00:50:39.500 | And how we can create something that mimics

00:50:41.500 | That resembles the fundamentals

00:50:44.500 | Of what makes our brain intelligent

00:50:46.500 | And finally, the Turing test

00:50:48.500 | The traditional definition of intelligence

00:50:51.500 | Defined by Alan Turing

00:50:52.500 | Was grounded in natural language processing

00:50:55.500 | Creating chatbots that impress us

00:50:57.500 | That amaze us

00:50:58.500 | And trick us into thinking they're human

00:51:00.500 | We will have a project

00:51:03.500 | And a speaker

00:51:04.500 | On natural language processing

00:51:06.500 | In March

00:51:08.500 | With that, I'd like to thank you for coming today

00:51:10.500 | And look forward to seeing your submissions

00:51:12.500 | For the three projects

00:51:14.500 | Thank you very much

00:51:15.500 | Thank you very much.

MIT AGI: Artificial General Intelligence

Chapters