This is the opening lecture for course 6.S099: Artificial General Intelligence. This class is free and open to everyone. Our goal is to take an engineering approach to exploring possible paths toward building human-level intelligence for a better world.

INFO:
Slides: http://bit.ly/2HdCm1N
Website: https://agi.mit.edu
GitHub: https://github.com/lexfridman/mit-deep-learning

CONNECT:
- If you enjoyed this video, please subscribe to this channel.
- AI Podcast: https://lexfridman.com/ai/
- Show your support: https://www.patreon.com/lexfridman
- LinkedIn: https://www.linkedin.com/in/lexfridman
- Twitter: https://twitter.com/lexfridman
- Facebook: https://www.facebook.com/lexfridman
- Instagram: https://www.instagram.com/lexfridman
- Slack: https://deep-mit-slack.herokuapp.com

Welcome to course 6S099 Artificial General Intelligence.

We will explore the nature of intelligence From as much as possible an engineering perspective.

What that means is we want to explore the fundamental science of what makes an intelligent system.

The core concepts behind our understanding of what is intelligence.

But we always want to ground it in the creation of intelligence systems.

We always want to be in the now, in today, in understanding how today we can build artificial intelligence systems that can make for a better world.

First and foremost, we're scientists and engineers.

We want to provide with this approach a balance to the very important but overrepresented view of artificial general intelligence.

Where the idea is once we know how to create a human level intelligence system, how will society be impacted?

Will we achieve a utopia that will remove the need to do any of the messy jobs that will make us all extremely happy?

Those kinds of beautiful philosophical concepts are interesting to explore, but that's not what we're interested in doing.

I believe that from an engineering perspective, we want to focus on the black box of AGI.

Start to build insights and intuitions about how we create systems that approach human level intelligence.

I believe we're very far away from creating anything resembling human level intelligence.

However, the dimension of the metric behind the word far may not be time.

In time, perhaps through a few breakthroughs, maybe even 1 breakthrough, everything can change.

But as we stand now, our current methods as we will explore from the various ideas and approaches and the guest speakers coming here over the next 2 weeks and beyond.

Our best understanding, our best intuition and insights are not yet at the level of reaching without a major leap and breakthrough and paradigm shift towards human level intelligence.

So it's not constructive to consider the impact of artificial intelligence, to consider questions of safety and ethics, fundamental, extremely important questions.

It's not constructive to consider those questions without also deeply considering the black box of the actual methods of artificial intelligence, human level artificial intelligence.

And that's what I see, what I hope this course can be.

Its first iteration, its first exploratory attempt to try to look at different approaches of how we can engineer intelligence.

It's to consider the big picture, the future impact of society 10, 20, 30, 40 years out, but fundamentally grounded in what kind of methods do we have today and what are their limitations and possibilities of achieving that, the black box of AGI.

And in the future impact on society of creating artificial intelligence systems that get become increasingly more intelligent.

The fundamental disagreement lies in the fact the very core of that black box, which is How hard is it to build an AGI system?

How hard is it to create a human level artificial intelligence system?

From Josh Tenenbaum to Andrej Karpathy to folks from OpenAI to Boston Dynamics to the brilliant leaders in various fields of artificial intelligence that will come here.

There's been a lot of incredibly impressive results in deep learning, in neuroscience, in computational cognitive science, in robotics.

But how far are we still to go to the AGI?

That's the fundamental question that we need to explore before we consider the questions, the future impact on society.

And the goal for this class is to build intuition, 1 talk at a time, a project at a time, Build intuition about where we stand, about what the limitations of current approaches are, how can we close the gap.

A nice meme that I caught on Twitter recently of the difference between the engineering approach at the very simplest of a Google intern typing a for loop that just does a grid search on parameters for neural network.

And on the right is the way media would report this for loop.

I think it's easy for us to go 1 way or the other, but we'd like to do both.

Our first goal is to avoid the pitfalls of black box thinking, of the futurism thinking that results in hype that's detached from scientific engineering understanding of what the actual systems are doing.

That's what some of our speakers will explore in a rigorous way.

It's still an important topic to explore.

Ray Kurzweil on Wednesday will explore this topic.

Next week, talking about AI safety and autonomous weapon systems, we'll explore this topic.

How do we design systems today that would lead to safe systems tomorrow?

But the reality is a lot of us need to put a lot more emphasis on the left, on the for loops, on creating these systems.

At the same time the second goal of what we're trying to do here is not emphasize the silliness, the simplicity, the naive basic nature of this for loop in the same way as was the process in creating nuclear weapons before during World War II.

The idea that as an engineer, as a scientist, that I'm just a scientist is also a flawed way of thinking.

We have to consider the big picture impact, the near term negative consequences that are preventable, the low hanging fruit that can be prevented Through that very engineering process We have to do both And in this engineering approach We always have to be cautious that just because we don't understand, just because we, our intuition, our best understanding of the capabilities of modern systems that learn, that act in this world, seem limited, seem far from human level intelligence.

Our ability to learn and represent common sense reasoning seems limited.

The exponential, potentially exponential, as could be argued, and he will, growth of technology, of these ideas, means that just around the corner is a singularity, is a breakthrough idea that will change everything.

Moreover, we have to be cautious of the fact that every decade over the past century, our adoption of new technologies has gotten faster and faster.

The rate at which a new technology from its birth to its wide mass adoption has shortened and shortened and shortened.

That means that new idea, the moment it drops into the world can have widespread effects overnight.

So as, and I think the engineering approach is fundamentally cynical on artificial general intelligence because every aspect of it is so difficult.

We have to always remember that overnight everything can change.

Through this question of beginning to approach from a deep learning perspective, from deep reinforcement learning, from brain simulation, computational cognitive science, from computational neuroscience, from cognitive architectures, from robotics, from legal perspectives and autonomous weapon systems, As we begin to approach these questions, we need to start to build intuition.

How far away are we from creating intelligent systems?

The singularity here is that spark, that moment when we're truly surprised by the intelligence of the systems we create.

I'd like to visualize it by the certain analogy that we're in this dark room looking for a light switch with no knowledge of where the light switch is.

There's going to be people that say, well, the rooms are all small, we're right there, anywhere we'll be able to find it in any time.

So we have to stumble around, feel our way around to build the intuition of how far away we really are.

And many speakers here will talk about how we define intelligence, how we can begin to see intelligence, what are the fundamental impacts of creating intelligence systems.

I'd like to sort of see the positive reason for this little class and for these efforts that have fascinated people throughout the century of trying to create intelligent systems is that there's something about human beings that craves to explore, to uncover the mysteries of the universe.

Fundamental in itself, a desire to uncover the mysteries of the universe.

Not for a purpose, and there's often an underlying purpose of money, of greed, of the power, craving for power and so on, but there's seems to be an underlying desire to explore.

Nice little book, An Exploration, a very short introduction by Stuart Weaver.

He says, for all the different forms it takes in different historical periods, for all the worthy and unworthy motives that lie behind it, Exploration, travel for the sake of discovery and adventure is a human compulsion, a human obsession even.

It is defining element of a distinctly human identity and it will never rest at any frontier, whether terrestrial or extraterrestrial.

From 325 BCE, with a long 7,500 mile journey on the ocean to explore the Arctic, to Christopher Columbus and his flawed, harshly criticized and modern scholarship trip that ultimately paved the way, didn't discover, paved the way to colonization of the Americas.

To the Darwin trip, the voyage of the Beagle, whilst this planet has gone cycling on according to the fixed law of gravity.

From so simple a beginning, endless forms, most beautiful and most wonderful have been and are being evolved.

To the first Venture into Space by Yuri Gagarin, first human in space in 1961.

What he said over the radio is the earth is blue.

These are the words that I think drive our exploration in the sciences, in the engineering, and today in AI.

From the first walk on the moon, and now the desire to colonize Mars and beyond.

That's where I see this desire to create intelligence systems.

Talking about the positive or negative impact of AI in society.

Talking about the business case of the jobs lost, jobs gained, jobs created.

The autonomous vehicles, the ethical questions, the safety of autonomous weapons, of the misuse of AI in the financial markets.

Underneath it all, and there are people, many people have spoken about this, what drives myself and many in the community is the desire to explore, to uncover the mystery of the universe.

And I hope that you join me in that very effort with the speakers that come here in the next 2 weeks and beyond.

The website for the course is agi.mit.edu.

I am a part of an amazing team, many of whom you know.

For registered MIT students, you create an account on the website and submit 5 new links and vote on 10 to vote AI, which is an aggregator of information and material we've put together for the topic of AGI.

And submit a entry to 1 of the competitions, 1 of the 3 competitions projects that we have in this course.

And the projects are Dream Vision, I'll go over them in a little bit.

Dream Vision, Angel, Ethical Car, and the aggregator of Material Vote AI.

We have guest speakers, incredible guest speakers, I will go over them today.

And as before, with the deep learning for self-driving cars course, we have shirts and they're free for in-person, for people that attend in person for the last lecture most likely, or you can order them online.

We take the Google Deep Dream idea, we explore the idea of creativity, or Einstein's view of intelligence, the mark of intelligence is creativity.

This idea is something we explore by using neural networks in interesting ways to visualize what the network see and in so doing create beautiful visualizations in time through video.

So taking the ideas of Deep Dream and combining them together with multiple video streams to mix dream and reality.

And the competition is through Mechanical Turk.

We set up a competition of who produces the most beautiful visualization.

We'll provide code to generate this visualization and ideas of how you can make it more and more beautiful and how to submit it to the competition.

Angel, the artificial neural generator of emotion and language is a different twist on the Turing test where we don't use words.

We only use emotions to speak, expression of those emotions.

And we create, we use a face, customizable, with 26 muscles, all of which can be controlled with an LSTM.

We use a neural network to train the generation of emotion.

And The competition in you submitting the code to the competition is you get 10 seconds to impress with these expressions of emotion the viewer.

Your goal is to impress the viewer enough to where they choose your agent versus another agent.

And those that are most loved, the agents most loved, will be the ones that are declared winners.

In a twist, we will add human beings into this mix.

So we've created a system that maps our human faces, myself and the TAs, to where we ourselves enter in the competition and try to convince you to keep us as your friend.

Ethical car, building on the ideas of the trolley problem and the moral machine done here in the media lab, the incredible interesting work.

We take a machine learning approach to it and take what we've developed, the deep reinforcement learning competition for 6S094, the deep traffic, and we add pedestrians into it.

Stochastic, irrational, unpredictable pedestrians.

And we add human life to the loss function where there's a trade-off between getting from point A to point B.

So in deep traffic, the deep reinforcement learning competition, the goal was to go as fast as possible.

Here, it's up to you to decide what your agent's goal is.

There's a Pareto front trade-off between getting from point A to point B as fast as possible and hurting pedestrians.

Because fundamentally in creating autonomous vehicles that function in this world, we want them to get from point A to point B as quickly as possible.

The United States government insurance companies put a price tag on human life.

We put that power in your hands in designing these agents to ask the question of how can we create machine learning systems where the objective function, the loss function has human life as part of it.

And Vote.ai is an aggregator of different links, different articles, papers, videos on the topic of artificial general intelligence where people vote on, vote quality articles up and down and choose on the sentiment of positive and negative.

We'd like to explore the different ways, the different arguments for and against artificial general intelligence.

There is an incredible list of speakers, the best in their disciplines.

From Josh Tenenbaum here at MIT, to Ray Kurzweil at Google, to Lisa Feldman Barrett and Nate Drabinski from Northeastern University, Andrej Karpathy, Stephen Wolfram, Richard Moyes, Mark Reiber, Ilyas Eskiver, and myself.

I'd like to go through each of these speakers and talk about the perspectives they bring.

To try to see the approach, the ideas they bring to the table.

They're not, in most cases, interested in the discussion of the future impact on society without grounding it into the expertise, into the actual engineering, into creating these intelligent systems.

So Josh is a computational cognitive science expert, professor, faculty here at MIT.

He will talk about how we can create common sense understanding, systems that see a world of physical objects and their interactions and our own possibilities to act and interact with others.

The intuitive physics, how do we build into systems the intuitive physics of the world more than just the deep learning memorization engines that take patterns and learn through supervised way to map those patterns to classification.

Actually begin to understand the intuitive, the common sense physics of the world.

And learn rapid model-based learning, learn from nothing, learn from very little.

Just like we do as children, just like we do as human beings successfully, often only need 1 example to learn a concept.

How do we create systems that learn from very few, sometimes a single example?

And integrate ideas from various disciplines, of course from neural networks, but also probabilistic generative models and symbol processing architectures.

Of course, from a different area of the world, another incredible thinker, intellectual speaker is Ray Kurzweil.

And he will do a whirlwind discussion of where we stand with intelligence, creating intelligence systems, how we see natural intelligence, our own human intelligence, how we define it, how we understand it, and how that transfers to the increasing exponential growth of development of artificial general intelligence.

Something I'm myself very excited about is Lisa Feldman Barrett coming here on Thursday.

She's written a book, I believe, How Emotions Are Made.

She argues that emotions are created, that there is a distinction, there is a detachment between what we feel in our bodies, the physical state of our bodies, and the expression of emotion from body to the contextually grounded, to the face expressing that emotion.

Which means, now why is there a person who is a psychology person in a fundamental engineering computer science topic like AGI?

Because if emotions are created in the way she argues and she'll systematically break it down, that means we're learning societal, as human beings we're learning societal norms of how to express emotion.

The idea of emotional intelligence is learned, which means we can have machines learn this idea.

It's a machine learn, just like it's a human learning problem, it's a machine learning problem.

In a little bit of a twist, she asked that instead of giving a talk, I have a conversation with her.

So it's going to be a little bit challenging and fun and she's great, I'm looking forward to it.

And we'll explore different ways that we can get emotion expressed through video, through audio, through the project, the Angel project that I mentioned.

So there's been work in reenacting intelligence.

So, well, reenacting, mapping face to face, mapping different emotions on video that was previously recorded.

So if you can imagine, that means we can take emotions that we've created, the kind of emotion creation we've been discussing, and remap it on previous video.

That's 1 way to see intelligence is taking raw human data that we already have and mapping new computer generated.

The underlying fundamentals are human but the surface appearance, the representation of emotion, visual or auditory, is generated by computer.

It could be in the embodied form like with Sophia.

I think it will be similar in a lot of ways, but different in a few others.

to develop complex emotions, and possibly robots can be

able to develop the more problematic emotions, like rage, jealousy, hatred, and so on.

It might be possible to make them more ethical and humans, so I think it would be a good partnership where 1 party completes the other.

Very important to note, for those captivated by Sophia in the press or have seen these videos, Sophia is an art exhibit.

She's not a strong natural language processing system.

But it's a beautiful visualization of embodying, of how, it's a beautiful visualization of how easy it is to trick us human beings that there's intelligence underlying something.

That the emotional expression, the physical embodiment and the emotional expression that has some degree of humor, that has some degree of wit and intelligence is enough to captivate us.

So that's an argument for not creating intelligence from scratch, but having machines at the very surface, the display of that emotion, the generation, the mapping of the visual and the auditory elements where underneath it is really trivial technology that's fundamentally relying on humans like in the Sophia's case.

And in the simplest form, we remove all elements of, how should I say, attractive appearance from an agent.

We really keep it to the simplest muscles, aspect characteristics of the face and see with 26 muscles controlled by neural network through time, so recurrent neural network, LSTM, how can we explore the generation of emotion?

Can we get this thing, and this is an open question for us too, we just created this system, we don't know if we can.

Make us feel something by watching it express its feelings?

Can it learn to by competing against other agents?

A, B, testing on Turk, on Mechanical Turk.

Can the winners be very convincing to make us feel entertained, pity, love?

Maybe some of you will fall in love with Angel here.

Nate Derbenski on Friday will talk about Cognitive Modeling Architectures.

So he will speak about the cognitive modeling aspect.

Can we have a, can we model cognition in some kind of systematic way to try to build intuition of how complicated cognition is.

Andrej Karpathy, famous for being the state-of-the-art human on the ImageNet challenge, the representative, the 95% accuracy performance, among other things he's also famous for, is now at Tesla.

He will talk about the role, the limitations, the possibilities of deep learning.

We'll talk as I have spoken about in the past few weeks and throughout about our misunderstanding or our flawed intuition about what are the difficult and what are the easy problems in deep learning.

And the power of representational learning, the ability of neural networks to form deeper and deeper representations of the underlying raw data that ultimately forms, that takes complex information that's hard to make sense of and convert it into useful actionable knowledge.

That is from a certain lens, in a certain lens, in a certain problem space can be clearly defined as understanding of the complex information.

Understanding is ultimately taking complex information and reducing it to its simple essential elements.

Representational learning is in the trivial case here in drawing, having to draw a straight line to separate the blue and the red curves.

That's impossible to do in a initial input space on the left.

What the act of learning is for deep neural networks in this formulation is to construct a topology under which there exists a straight line to accurately classify blue versus red.

And for a simple blue and red line, it seems trivial here, but this works in a general case for arbitrary input spaces, for arbitrary nonlinear highly dimensional input spaces.

And the ability to automatically learn features, to learn hierarchical representations of the raw sensory data means that you could do a lot more with data, which means you can expand further and further and further to create intelligence systems that operate successfully with real world data.

That's what representation learning means.

That deep learning allows, because the arbitrary number of features that can be automatically determined, you can learn a lot of things about a pretty complex world.

Unfortunately, there needs to be a lot of supervised data.

There still needs to be a lot of human input.

Andre and others, Josh, will talk about the difference between our human brain, our biological neural network and the artificial neural network.

The full human brain with 100 billion neurons, 1,000 trillion synapses and the biggest neural networks out there, the artificial neural networks, having much smaller, 60 million synapses for ResNet-152.

The biggest difference, the parameters of human brain being several orders of magnitude more synapses.

The topology being much more complex, chaotic.

The asynchronous nature of the human brain.

And the learning algorithm of artificial neural networks is trivial and constrained with back propagation.

Is essentially an optimization function over a clearly defined loss function from the output to the input.

Using back propagation to teach, to adjust the weights on that network.

The learning algorithm for our human brain is mostly unknown But it's certainly much more complicated than back propagation The power consumption, the human brain is a lot more efficient Than artificial neural networks And there's a very kind of artificial, trivial supervised learning process for training artificial neural networks.

You have to have a training stage and you have to have an evaluation stage.

And once the network is trained, there's no clear way to continue training it or there's a lot of ways but they're inefficient.

It's not designed to do online learning naturally, to always be learning.

It's designed to be to learn and then be applied.

Obviously our human brains are always learning.

But the beautiful fascinating thing is that they're both distributed computation systems on a large scale.

So it's not a, It doesn't ultimately boil down to a single compute unit.

The back propagation learning process is distributed.

Can be paralyzed in a GPU, massively paralyzed.

The underlying computational unit of a neuron is trivial But can be stacked together to form forward neural networks Recurrent neural networks To represent both spatial information with images And temporal information with audio speech Text, sequences of images and video And so on Mapping from 1 to 1, 1 to many, many to 1.

So mapping any kind of structure, vector, and time, data as an input to any kind of classification, regression, sequences, captioning, video, audio as output.

Learning in the general sense, but in a domain that's precisely defined for the supervised training process.

We can think of the, in deep learning case, you can think of the supervised methods where humans have to annotate the data as memorization of the data.

See all Lex Fridman transcripts on Youtube

MIT AGI: Artificial General Intelligence