This is a guest talk part of MIT 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:
Course website: https://agi.mit.edu
AI podcast: https://lexfridman.com/ai

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

Today we're going to try something different.

She is a university distinguished professor of psychology at Northeastern University, director of the Interdisciplinary Effective Science Laboratory author of the new amazing book, How Emotions Are Made, The Secret Life of the Brain.

She studies emotion, human emotion, from social, psychological, cognitive science and neuroscience perspectives.

And so I think our conversation may help us gain intuition about how we instill emotional life into future artificial intelligence systems.

As Josh Tenenbaum gave you a shout out on Tuesday and said that if you want to understand how to create artificial general intelligence systems from an engineering perspective, we should study the human brain.

Let me start with a curve ball to conjure up an image of emotion.

Have you ever cried while watching a movie that you remember, and what movie was it?

The last time I cried, actually, here's an interesting thing.

When I'm speaking about, when I'm giving an academic talk, I sometimes will talk about a study that we did where we had people watch films, and we have them watch the most evocative clips of films.

And there's several clips that are really powerful, and a couple of them, every time I talk about them, I'm gonna try not to cry now, every time I talk about them, I'm describing for the audience what subjects are seeing.

1 of the clips is from a movie called Sophie's Choice.

So we show, this is a film about a woman who is forced in a concentration camp to choose which of her children will die in the gas chambers.

And so I'm already, you know, if I had a heart rate monitor and respiration monitor, you would see I'm, it's a really powerful scene.

Meryl Streep is Sophie, and it's very, very evocative.

We also show, there's a scene from a film with Susan Sarandon who is dying of breast cancer and she has to tell her 12 year old daughter that she's dying.

So it's another scene also that I find very compelling.

And you use these scenes to elicit emotion as part of experiments.

And in fact I was just giving a presentation to the Supreme Judicial Court of Massachusetts on implicit bias and to start the, you know, they wanted to understand the neuroscience of implicit bias and whether or not they should be crafting jury instructions for juries to be aware of implicit bias.

And to open the discussion, I showed them a clip from a film that was filmed almost 20 years ago actually, called A Time to Kill, a scene where Matthew McConaughey is this kind of, it's his closing statements in a case where he is defending an African American man who murdered 2 European American men who raped his 12 year old daughter.

And so the scene is, he's this completely pathetic lawyer until the end when he masters this fantastic defense basically of this man on trial for you know basically avenging his daughter's death and that 1 I hadn't seen that for 20 years, that film, and it brought me to tears, actually.

It's just a really powerful, powerful scene, and it's, it just really punches you in the stomach.

You just can't help but just experience the weight of racism in that room and the brilliance of the closing argument to sort of puncture through it basically.

So 1 of the things you talk a lot in your work, in your book, is the difference in the experience of emotion, the expression of emotion.

So what's our biggest misconception about emotion for folks that haven't considered, have only considered emotions at a surface level.

I would say 1 common misconception is that you can look at someone's face and read their emotion the way you read words on a page.

That everyone around the world, when they're feeling angry, they scowl.

When they're feeling sad, that they frown.

The way that I'm saying it sounds preposterous, and it is preposterous, but unfortunately, that is actually what a lot of people believe.

And tech companies around the world spend tremendous amounts of money.

And you know, in the investment of some of the most creative people on this planet, trying to build emotion detection systems, when really what they're building are excellent systems for reading facial movements which have no intrinsic emotional meaning.

I mean sometimes facial movements communicate emotion but many times they don't.

They don't just smile when they're happy.

They sometimes smile to indicate, to send a social message that has nothing to do with emotion.

And so the idea that there's 1 universal facial expression, for example, for each emotion is just, there's no strong scientific evidence for that claim.

Because we nevertheless feel like we're observing emotion when we're communicating with others.

So how does emotion, the creation of emotion change in the presence of others?

Does the audience change the display of emotion?

Well, emotions are not displayed, I would say, right?

So basically, you and I are having a conversation right now and part of what my brain is doing is it's guessing, making educated guesses about what your facial movements mean.

So right now, you have a slight smile on your face, not exactly a smirk, but.

the way, so we're not allowed to show any emotion.

You know, I have a friend of mine who is Russian who told me, and when she moved to this country, actually I've had several friends tell me this, that their cheeks hurt for a year for how much smiling they had to do.

I have also a friend from the Netherlands who moved here who told me the same thing.

Her face ached for how much smiling she did.

And in Bulgaria they have an apparently like a name for you know the pervasive American smile, kind of like a not flattering name for how much Americans smile.

But basically when you look at someone's face you're making a guess about how they feel.

And although you yourself are focused on their face, to you it feels as if that's where all the information is.

Your brain is actually taking in the entire sensory array.

So it's taking in also, you know, the sound of the person's voice and the person's body posture and the dynamic temporal changes in all of those signals to make sense of things.

It's not, you're not really, the face is not a lighthouse, it's not a beacon that just displays somebody's internal state in an obligatory way.

And so what are, what would you say as far as I understand there's no good answer yet from science, is what are the basic building blocks of emotion?

Well, I wouldn't say that there's no good answer from science, I would say scientists disagree.

I think it's very clear what the building blocks are.

But, you know, in 1 way or another, if you look back all the way to ancient Greece, you can see that there are 2 kinds of views of emotion that have been battling it out for millennia.

1 is the idea that you have, I will give you the modern versions of these views, but they really do have a very long history.

1 is the idea that you are born with circuits in your brain that are pre-wired for emotion.

So that everybody around the world is born with an anger circuit, a fear circuit, a sadness circuit, a happiness circuit, some other circuits too.

You have them, we all have them, all neurotypical brains have them.

And actually other animals have them too.

And the idea is that when 1 of these circuits is triggered you have an emotional response which is obligatory.

So you have a very stereotypic change in your body, your breathing changes in a stereotypic way, your heart rate changes, the chemicals in your body change in a particular way, you make a particular facial expression, and you have a propensity to make a particular action, like you attack in anger and you run in fear, or you freeze in fear.

And then there's another view which says, well, there are some basic kind of ingredients or building blocks that the human mind or the human brain now, people talk about the brain now, there's some basic ingredients that your brain uses and it uses them to make every kind of mental event that you experience and every action that you take.

Just like you can take flour and water and salt and you can make a whole bunch of different recipes with them, some of which aren't even food, like glue.

In a similar way, the idea is that your brain has some kind of all-purpose capacities and it puts these ingredients together and it makes emotion as you need them on the spot.

And you don't have 1 anger, you have a whole repertoire of anger.

You don't have 1 sadness, you have a whole repertoire of sadness.

And if you grow up in a culture that has no concept for sadness, you don't experience sadness and you don't perceive sadness because your brain becomes wired to make whatever mental events that exist in your particular culture.

So for artificial intelligence systems, the idea of emotional intelligence is really difficult and it seems like it's an important thing to try to instill.

So for human beings, where do you see the importance on the priority list of what makes us human?

I actually think that's the wrong question to ask because, no, I mean, I think people constantly ask that question, but I don't think it's the right question to ask because some cultures on our planet don't even have a concept for emotion, and people don't, while they experience what I'm going to refer to as, in scientific terms, we call affect, which is our simple feelings of feeling pleasant or unpleasant or feeling worked up or calm or comfortable or just or you know having discomfort.

These are feelings that come directly from the internal state of the physical systems in your body, not everybody makes emotions out of those feelings.

And in fact, we often don't make emotions out of those.

So those feelings, the way to think about it is that your brain comes wired to regulate your body.

Your brain comes wired, when you're born your brain comes wired with the potential to regulate your body.

Infants actually don't regulate their own nervous systems very well.

They can't regulate their own temperature.

And as caregivers regulate the nervous systems of their infants, that wires the infant's brain.

So infant, little infant brains aren't born like little miniature adult brains.

They're born waiting for a set of wiring instructions from the world and they wire themselves to the physical and social realities that they grow in And as they learn to do this, they experience these simple feelings that feeling pleasant or unpleasant, feeling worked up or feeling calm, these are kind of like barometer readings in a way.

They're very simple and they lack a lot of detail because of how we're wired.

And the way that a culture makes sense of them is not always about emotion.

So I would say, in our culture, when we lose something significant like a loved 1, we feel sad.

100 or 200 years ago, there was an emotion called nostalgia which killed people, was thought to kill people after, you know, for example, after serving in World War I, we would now call that depression.

It's not just a matter of changing the label of something, it's actually the formation of the experience is very different.

So if you want to build an intelligent agent, I don't think emotion is what you need to endow it with.

You need to endow it with these basic ingredients that it can use to make whatever experiences or guide whatever actions, make whatever states or guide whatever actions are relevant to the situations that it's in.

That will be different if it's an American or a British or a Western urbanized environment than say if you were to go to Tanzania and study you know the Hadza who are hunting and gathering since that culture since the Pleistocene.

So do you have words or human interpretable labels on these basic ingredients that we can understand?

I mean I think the first thing to understand is that when we think about building an intelligent agent we think about endowing the agent with cognition or with emotion or with the ability to perceive things in the world because as humans, these are the things that are really important to us especially in our culture, thinking, feeling, seeing.

In other cultures, you know, they have different ways of parsing things.

But the truth is that your brain did not evolve so that you could think and feel and see.

It evolved, brains evolved to control bodies.

Brains evolved so that they could control the various systems of the body as creatures move around in order to gain resources like food and water.

And a brain has to figure out how much resources to expend on getting additional resources.

Now that may sound really trivial, but it turns out it's actually really hard.

And scientists actually haven't really figured out completely how brains do this.

I mean what's really interesting to me is that if you look for example at computer vision, try to figure out you know how and how to make an agent see, that's a pretty much solved problem.

Not completely, it's a pretty much solved problem.

However, if you want to create an agent that just reaches out smoothly, grabs a glass, and brings it into the body to drink, that problem is not solved.

Something as simple as movement, which we think of as this really basic thing like, oh, it's so trivial, all animals can do it, is actually 1 of the hardest problems to solve.

And brains basically, I mean, there's a whole story here about evolution which has nothing to do with having a lizard brain which is wrapped in a cognitive brain or anything like that.

That's a reference to the triune brain which a lot of people believe.

It's a way of thinking about brain evolution.

Instead, what seems to be the case is that as bodies got larger and there were more systems because the niche of the animal, the environment of the animal got bigger and bigger and bigger, brains also had to get bigger.

But they had to get bigger to a point with some constraints.

Like they have to be, you have to keep metabolic costs down.

If you don't, creatures get sick and die.

We can talk about what that means in terms of depression or metabolic illnesses or what have you.

Well, 1 of the basic ingredients is that your brain is controlling your body all the time.

Whether you feel it or not, whether you're thinking about it or not, whether you're asleep or awake, certain parts of your brain are always very active all the time, even when you're sleeping or else you'd be dead.

And those parts of the brain that are controlling your heart and your lungs and your immune system and all of those regions that are controlling the systems of your body are also helping your brain to represent the sensory consequences of those changes in your body which you don't feel directly.

You don't feel your heart beating most of the time.

You don't feel your lungs expanding most of the time.

And there's a really good reason why we are all wired not to feel those things.

And that is you'd never pay attention to anything outside in the world ever again if you could feel every little movement that was going on inside your body.

So your brain represents those as a summary, these kind of simple summary feelings.

You feel really jittery, you feel really calm.

They sometimes, your brain can make them into emotions, but they're with you every waking moment of your life.

They are properties of consciousness in the way that lightness and darkness is a property of vision.

And sometimes we make them into emotions.

When we have a big change in our heart rate or a big change in our breathing rate or a big change in our temperature or a big surge of glucose, the brain might make a motion out of those changes, those very strong changes which you will feel as really feeling unpleasant or really feeling pleasant.

But your brain might also make hunger, or your brain might make an instance of a physical sensation like nausea.

Or your brain might even make a perception of the world, like, that's a nice guy.

Those are also moments where these simple feelings, which we call affect, are very strong.

There are others that I could talk about too, it's not the only 1, but 1 of the things I think that sometimes people who are studying to build AI systems don't realize is that the brain, its fundamental job is to keep your body alive and well.

And if you don't have some kind of body to regulate with affective feelings that come from that regulation or something like that, you're kind of going to be out of luck, I think, in rendering something that looks more, something that seems more human.

Right, so maybe you can elaborate, like in the book Sapiens, that as human beings, we're really good en masse as with thousands, millions of people together believing something even if it's not true.

So, while scientifically sort of from a neuroscience perspective it's maybe very true that emotions aren't real.

I didn't say they aren't real, but I said they're not.

There isn't, so it's interesting, well, you finish and

So what I'm trying to say is, also from an AI perspective, is they become these trivial ideas of mapping a smile to being happy and these kind of trivial ideas become real to us through Hollywood, through cultural spreading of information and we start to believe this and therefore it becomes real.

In as much as anything is real about our perception together.

So it's really important scientifically, the ideas that you're presenting, but does that mean there's, just because our brain doesn't feel those explicit emotions, does that mean they're not real?

I didn't say we don't feel explicit emotions.

So I want to be really clear about this because it's an interesting, this inference, it's an interesting inference that people often make.

And it's a mistake that betrays a certain kind of thinking that we do in this culture.

And it's the mistake of the following sort.

So when I say there's no facial expression that is diagnostic of a single emotion.

That doesn't mean that people don't express emotion.

They just don't, you know, when they're in a state of anger, in a state of sadness, or in a state of awe, their faces don't do 1 thing.

When I say, well your body can do many things when you're angry or when you're sad or but let's take anger.

Your heart rate can go up, it can go down, it can stay the same.

Your breathing rate can go up, it can go down, it can stay the same.

The same pattern that you see in anger you sometimes see in sadness and you sometimes see in fear.

You sometimes even see it in enthusiasm and in awe.

So does that mean that emotions aren't real?

No, emotions are real, But sometimes things are real because the physical meaning of the signal is endowed in the signal, okay?

So when your retina communicates to your brain that you are faced with a wavelength that is 600 nanometers, the signal is endowed in your brain.

You don't interpret that it's 600 nanometers.

But when you see red, that information is not in the signal.

Your brain has added information that isn't in the signal itself.

In a sense, your brain has imposed meaning on a signal that the signal doesn't have on its own.

Now let me back up and give a different example to make it a little easier, and then we'll re-approach this.

There's a, There are some things that we, in fact, this is true almost of all civilization, right?

That we, there are some things that are real by virtue of the fact that we agree that they exist.

Little pieces of paper serve as money or now you know Bitcoin or little pieces of plastic or gold or diamonds or in the past barley, salt, shells, rocks, serve as currency, have value, only because a group of people agree that they have value.

So we impose meaning on objects, and once we all agree that that object actually has value, we can trade it for material goods.

The minute that some of us disagree, that we withdraw our consent, the things lose their value.

That's what happened in the mortgage crisis.

That's what happened in the tulip crisis in the Netherlands in the 17th century or 16th century, or whatever it was.

Money, currency exists because we impose meaning on objects in the world, physical objects in the world that themselves don't have that meaning on their own.

And they are very real, money is very real to people.

I can stick somebody's head in a brain scanner and show you that they experience value in a very real way.

But that reality is constructed by the fact that they have learned the value in a particular culture.

Well, that's also what we do with emotion.

We impose meaning on certain physical signals that they don't have on their own.

And But we have collective intentionality.

We all agree that scowling is sometimes anger, and so it becomes anger in a very real way, just like little pieces of paper become money.

So it sounds like you kind of think about the expression of emotion as a kind of language, as an extension of a language that we learn in the same way that we collectively agree on a language, on a lexicon and how we use that language.

I mean, Everything in our culture is, almost everything in our culture is a function of social reality in this way.

We are citizens of a country because we all agree that the country exists more or less.

And wow, you guys barely even laughed at that, okay.

A revolution is when some people in the country withdraw their consent.

A president has powers in a country because we all agree that a president has powers.

President only has powers by virtue of the fact that we all agree that he or she has powers.

If we stop agreeing, the president doesn't have those powers anymore.

People's lives depend, outcomes of real people depend on these social realities that we build and nurture.

And we wire these social realities into the brains of our children as we socialize them.

And when people move from 1 culture to another, they have to learn the new social reality that they are faced with.

And if they don't, they get very sick physically because our ability to agree on what something means actually is important for regulating our nervous systems.

So can you speak to that a little bit, I mean for machine learning methods, for systems that learn to behave based on a certain reward, it's important to kind of have some ground truth and learn.

So how do, it sounds like the expression of emotion is learned.

Can you talk about how we learn to fit into our culture by expressing emotion with our face, body, given the context, given the rich, what's that process look like?

Sure, well, you know, I mean, I wrote a 400 page book, so I'll try to do it in like a couple sentences.

So an infant's born and it can't really do anything for itself.

It can't regulate its own nervous system.

This is a term, scientific term for this is allostasis.

Allostasis is your brain's ability to predict what your body's gonna need before it needs it and tries to meet those needs before they arise.

So an example would be, if you're gonna stand up, if your brain's gonna stand you up, it has to raise your blood pressure before it stands you up.

That's costly from a metabolic standpoint.

So an infant's brain can't do this very well.

An infant doesn't know when to go to sleep and when to wake up.

An infant doesn't, can't feed itself, can't regulate its own temperature.

And when someone does it, the infant is learning.

It's taking in sights and sounds and smells and the physical sensations from the body which are comfortable and pleasant when the infant allostasis is maintained.

So right from the get-go, an infant is learning, statistical learning, capturing events, including their consequence for the infant's body.

Some people think babies are born, you know, attached already to their caregivers, but they're not actually.

Infants don't even know what a caregiver is.

It's just that the caregiver is there constantly meeting that infant's needs.

Now, if there are statistical regularities, like for example, an infant is not born with the ability to recognize a face as a face, but it learns that in like the first couple of days of life.

See all Lex Fridman transcripts on Youtube

Lisa Feldman Barrett: How the Brain Creates Emotions | MIT Artificial General Intelligence (AGI)