How the Child's Mind Informs AI Research - Alison Gopnik at BrainMind

Summary

The speaker, a philosopher turned neuroscientist, discusses the profound question of how we understand the world despite receiving limited sensory input. They highlight children's remarkable ability to learn and make sense of the world with minimal data, contrasting it with current AI's reliance on vast datasets. The talk explores the potential for AI to mimic children's model-building, curiosity, and social learning, suggesting these could lead to more efficient and adaptable machine learning.

Takeaways

• 🎓 The speaker's eclectic research interests stem from a foundational problem: understanding how we gain knowledge about the world despite receiving only a narrow stream of sensory input.
• 👶 Studying children is crucial because they actively learn about the world around them, which can provide insights into how humans acquire knowledge.
• 🧠 The speaker's work in developmental psychology has challenged previous assumptions about children's understanding of other people's minds, revealing that even infants are capable of such understanding.
• 🤖 Current AI systems are compared to children with overbearing parents, as they are heavily directed and lack the autonomy to explore and learn from their environment.
• 🧪 Experiments have shown that children as young as 18 months can infer intentions and help others, indicating a complex understanding of social dynamics.
• 🌟 Studying how children learn could help design AI systems that are more capable of generalizing knowledge and learning from limited data.
• 🧠 The speaker highlights three areas where children excel over current AI: model building, curiosity and exploration, and social learning.
• 🤖 AI systems could potentially be designed to be more curious, intrinsically motivated, and capable of learning from social interactions.
• 🔍 There's ongoing research into reinforcement learning where AI systems are rewarded not only for success but also for finding surprising or unexpected outcomes.
• 🌐 The trade-off between exploitation (using current knowledge effectively) and exploration (searching for new knowledge) is a key concept in AI and human cognition.
• 💭 The concept of consciousness is complex and varies greatly, suggesting that it may not be adequately captured by a single definition or theory.

Q & A

What is the foundational problem that the speaker has been concerned with throughout their career?

The foundational problem the speaker is concerned with is understanding how we know as much as we do about the world around us, given the limited sensory input we receive.

Why does the speaker believe studying children is a good way to answer questions about knowledge acquisition?

The speaker believes studying children is beneficial because they are actively acquiring knowledge about the world, similar to how AI systems are designed to learn from data.

What did the speaker and other developmental psychologists discover about children's understanding of other people's minds?

They discovered that even young babies are trying to figure out what's happening in other people's minds, contrary to previous beliefs that children are egocentric.

How did Felix Warneken's experiment with 18-month-olds demonstrate an understanding of others' intentions?

The experiment showed that 18-month-olds would give a dropped pencil to an adult, but not if the adult threw it, indicating they inferred the adult's intentions and were actively trying to fulfill a perceived desire.

What are the three things that children do, which current AI systems are not good at, according to the speaker?

The three things are: model building, being curious and exploratory, and learning socially from other people.

What is the significance of the work by colleagues at Berkeley mentioned by the speaker?

The work involves designing AI systems that are reinforced when they fail, encouraging them to explore and find surprising things, which leads to more robust learning compared to systems that only follow rewards.

What is the 'explorer exploit trade-off' mentioned in the script?

It refers to the intrinsic balance between being efficient at solving a problem and exploring the space of possibilities for potential solutions.

How does the speaker view the question of consciousness?

The speaker suggests that the question of consciousness might be a 'bad question', implying that it may be too complex or ill-defined to have a single answer, and that consciousness could be experienced in many different forms.

What does the speaker imply about the consciousness of babies compared to adult philosophers?

The speaker implies that babies' consciousness is very different from that of adult philosophers, suggesting that the latter's self-reflective consciousness might be an alteration from a more baseline state.

What does the speaker mean when they say consciousness could be a 'bad question'?

The speaker implies that the concept of consciousness is so complex and multifaceted that it may not be possible to define it in a way that encompasses all its forms and experiences.

Outlines

00:00

🔍 The Roots of Knowledge and Learning

The speaker, a philosopher turned neuroscientist, discusses their foundational interest in understanding how humans acquire knowledge about the world. They emphasize the narrow sensory input we receive and the vast knowledge we possess, questioning how this is possible. The speaker's focus is on children, who they see as the epitome of learning, as they actively interpret sensory data to understand the world. This leads to the idea of creating computers that can learn in a similar way, which is highly relevant in the age of AI. The speaker also touches on the importance of studying children's understanding of others' minds, which has been a significant area of research since the 70s and 80s, revealing that even infants are capable of understanding others' mental states.

05:02

🤖 Learning from Babies: Principles for AI

The speaker compares the learning abilities of children to current AI systems, noting that children can learn from very limited and messy data, forming generalizable principles that apply in various situations. This is in contrast to AI systems that require vast amounts of curated data. The speaker identifies three key areas where children excel and current AI struggles: model building, curiosity and exploration, and social learning. They suggest that to improve AI, we should consider incorporating these aspects, such as building AI systems that are intrinsically motivated, can learn from others, and can explain things rather than just predict outcomes. The speaker also mentions the work of colleagues who are designing machine learning algorithms that are reinforced when they fail, encouraging exploration and leading to more robust learning.

Keywords

💡Neuroscience

Neuroscience is the scientific study of the nervous system, including the brain. It encompasses the understanding of the nervous system's structure, function, development, genetic makeup, and the impact of disease and toxins. In the video, neuroscience is central to understanding how the brain processes information and gives rise to behaviors and cognitive processes such as perception, memory, and language.

💡Philosophy

Philosophy is the study of general and fundamental questions about existence, knowledge, values, reason, mind, and language. The speaker began their career in philosophy, which allowed them to ponder a wide array of subjects. Philosophy's influence is evident in the speaker's foundational problem of understanding how we come to know the world around us, a question that straddles both philosophical inquiry and empirical research.

💡Eclectic

Eclectic refers to a broad and diverse range of ideas or interests. The speaker mentions having an 'eclectic set of interests and influences,' indicating a multidisciplinary approach to their research, drawing from various fields to address complex questions about cognition and learning.

💡Developmental Psychology

Developmental Psychology is the scientific study of how humans change and develop throughout their lifespan. The speaker mentions that their work and that of other developmental psychologists in the 70s and 80s led to new techniques for understanding how children learn about the world and other people's minds.

💡Cognition

Cognition refers to the mental action or process of acquiring knowledge and understanding through thought, experience, and the senses. The script discusses how children's cognition is a critical area of study to understand how they make sense of the world, which is also relevant for designing AI systems that learn and understand.

💡Perception

Perception is the organization, identification, and interpretation of sensory information to represent the outside world. The speaker uses the example of how we get a 'narrow little stream of photons at the back of our eyes' to illustrate the process of perception, which is fundamental to understanding how we come to know about the world.

💡AI Spring

AI Spring refers to a period of rapid advancements in artificial intelligence. The speaker mentions the 'great new AI spring' in relation to computers that can learn from data, indicating a time of significant progress and interest in AI capabilities.

💡Machine Learning

Machine Learning is a subset of AI that provides systems the ability to learn and improve from experience without being explicitly programmed. The speaker discusses machine learning in the context of AI systems that can learn from data to make predictions or improve their performance.

💡Curiosity

Curiosity is the desire to know or learn something. The script highlights curiosity as a key aspect of how children learn, suggesting that AI could benefit from being designed to be intrinsically motivated and curious, similar to how children explore their environment.

💡Social Learning

Social Learning is learning that occurs within a social context, involving the transmission of knowledge, skills, values, and behaviors through observation and imitation. The speaker points out that babies learn socially, imitating and learning from others, which is a sophisticated process that could be incorporated into AI to improve its learning capabilities.

💡Consciousness

Consciousness is the state of being aware of and able to think and perceive. The speaker reflects on the concept of consciousness, suggesting that it may be a 'bad question' in the sense that it's too broad and may not have a singular answer. They also hint at the idea that there are different forms of consciousness beyond the self-reflective kind often considered by philosophers.

Highlights

Philosophy background influences neuroscience research

Interest in how we know about the world around us

Children's learning as a model for understanding cognition

Children actively figure out how the world works

Influence of developmental psychology on understanding children's knowledge

Children's understanding of other people's minds

Experiments on children's motivations and actions

Children's ability to infer intentions and help others

Studying children's learning to improve AI design

Current AI compared to children with helicopter parents

Children's learning principles are more generalizable

Three things children do that current AIs do not: model building, curiosity, social learning

AI's lack of intrinsic motivation and exploration

Children's sophisticated social learning and imitation

Designing AIs that are curious and intrinsically motivated

Reinforcement learning and the explorer-exploit trade-off

AI systems designed to be surprised and explore

Philosophical view on the question of consciousness

Consciousness as a spectrum with different states

Transcripts

i think of all the neuroscience

researchers i know you have one of the

most wide-ranging

and eclectic set of interests and

influences what influences

pre-neuroscience

do you think made you such have you have

such eclectic research interests well i

began my career and still am an

affiliate in the philosophy department

and one of the wonderful things about

being a philosopher is that you get to

think about all sorts of things but i do

think that even though on the surface

the interests that i have look

wide-ranging

there's a basic foundational problem

that's always been for my whole career

the problem that i'm concerned about and

that problem is how is it that we know

as much as we do about the world around

us

so if you look at the world we get a

narrow little stream of photons at the

back of our eyes and yet somehow we end

up knowing about a world full of people

and objects and places and science and

abstract things and the big question for

me has always been how does that happen

how is that possible it seemed to me

that a very good way of answering that

question was to look at children because

they're the ones who are doing that more

than any other creatures that we know of

they're the ones who are actually taking

that data and figuring out how the world

works and then that also leads to the

question of trying to understand what's

going on in their minds

what their motivations what their brains

and minds are like that enables them to

do that so effectively once that's what

you care about you can think about how

could you construct a computer that

could do the same thing and that's

become very relevant because the great

new ai spring has been about computers

that can learn that can take data and

make sense out of it you said that

scientists at the time who were trying

to understand how we come to understand

the world one didn't think that there

was any point in looking at children

they act as many physicists or

biologists and are actually able to come

up with

explanatory theories that they can then

use to make predictions about the world

my work and the work of a bunch of other

developmental psychologists starting in

the 70s and 80s

really basically found new techniques

for asking the question about what it is

the children knew first we discovered

that children understand things about

other people's minds

children were supposed to be so cystic

and egocentric and starting in the late

80s a number of developmental

psychologists started saying is that

really true

what do children understand about what's

going on in other people's minds and we

discovered that even even young babies

are trying to figure out what's

happening in other people's minds so

were there ways that we could ask them

in their language instead of our

language what they know and when we did

that

the sort of things that we would do is

look at what they do look at how they

act to try and help someone else so

the experiment you were talking about

very clever experiment by felix

wernicken who's now at

university of michigan he showed that

if you

took even an 18 month old and you

dropped a pencil on the floor the 18

month olds would come and give it to you

but not if you threw it to the floor

which means that they were both

inferring something about what you

wanted and also

actively trying to get you something

that you want what are some of the other

ways in which studying how

zero to ten year olds learn could help

us design ais better right so

one way that i like to put it now is if

you think look at our current ais

they're a bit like children who have

super hyper telecom helicopter tiger

moms so they have programmers who are

saying here's your score get your score

higher

and the great discovery of machine

learning in the last 10 years has been

you don't actually have to say get your

score higher by doing this you just tell

them get your score higher give them a

bunch of statistical data and billions

of examples they can figure out how to

do it themselves in some ways babies are

like the opposite of that so with very

small amounts of data

very messy data not well curated data

they seem to be able to learn

principles that are

much more generalizable that they can

apply in many more different

circumstances so the puzzle is what is

it that they're doing that's letting

them letting them

do that in a way that current ais can't

the three things that children but we

know babies are doing that current ais

are not very good at doing our

model building so actually building this

goes back to the work

i did in the 80s building theories ideas

about how the world works explaining how

the world works not just

predicting um they're curious they're

exploratory so the ais are kind of stuck

inside of their mainframes we can feed

them data but they can't go out and

get data for themselves the third thing

is learning socially so

babies are learning from other people

and they're extremely tuned into what

other people are doing they imitate

other people they listen to what other

people say but they don't just do this

in a kind of simple mindless way they do

it in a very sophisticated way so we're

taking some basic problems like figuring

out how objects work or how people work

and then trying to see could we get an

ai that is curious is intrinsically

motivated could we get an ai that can

learn from other people can imitate them

and imitate intelligently can we get an

ai that explains things that tries to

make up models and it doesn't just uh

doesn't just predict things the

curiosity

do you imagine uh

that a machine learning algorithm could

actually tell its humans

feed me a different kind of corpus of

information

there's beautiful work that my

colleagues at berkeley um poked agarol

and deepak paltech and

uh trevor daryl and anders andre malik

and ayosh efris have done

um of actually designing so one of the

big techniques in current machine

learning is reinforcement learning so

that's what i was saying you know you

have an alpha go it gets a score on an

atari game let's say and then it tries

to it gets reinforced for getting a

higher score and then it tries to figure

out how to get more reinforcement um

the very clever design that they have is

a system that is trying to build a model

of the world like kind of like

predictive coding trying to predict the

world but it also gets reinforced when

it fails

so it's going around essentially trying

to find things that are surprising

trying to find things that don't fit

with the way the world works and it gets

a little as it were shot of you know ai

dopamine when

when it's surprised when things are

weird when things don't work and it

turns out that that kind of a system

first of all does really explore the

space and it does it in a more robust

way than

uh the helicopter

ai that is just you know following the

trail of breadcrumbs of the rewards so

one of the really basic ideas that's in

ai and computation in lots of areas

is something that's called an explorer

exploit trade-off and the phenomenon is

there's an intrinsic trade-off between

what you need to do to be most effective

most efficient

get something done quickly and

effectively and what you need to do to

explore the space of possibilities and

you can kind of you know think see why

that's true right i mean there's big

giant space of possibilities only one of

them is actually or a small set is

actually going to be the one that's

going to work how do you do that how do

you explore possibilities

while you're exploring them you're not

actually effectively solving the problem

but when you're solving the problem

you're not exploring all the other ways

that you could solve the problem so

there's been some really interesting

work trying to see what humans do and

adult humans go back and forth between

different kinds of ways of trying to

solve the problem but it's really

challenging so finally um you came from

philosophy and then went into brain

science what what is consciousness

so i think that's one of those ones

where the answer is that it's a bad

question that philosophers this is an

old philosopher trick right

but i do think that's a that's true in

this case i think what's

who knows what the answer is going to be

but i think it's very unlikely that

there's going to be one answer perhaps

not coincidentally um the people who

have thought about consciousness tend to

focus on the kind of consciousness you

have when you're an adult philosopher

sitting in your armchair thinking about

consciousness which kind of makes sense

but that's really different from the

consciousness that you have when you're

a baby looking around in the world very

different from the kind of i think the

the recent work on psychedelics has

given us a lovely kind of um

demonstration of the fact that you can

have a kind of consciousness that's very

different from that sort of professorial

self-reflective consciousness you could

have a consciousness the characteristic

of which is that you don't feel a

difference between yourself and the

world anymore right when you start

casting your net more widely thinking

about animals thinking about children

thinking about

what people call altered states although

i think actually those maybe the sort of

baseline states from which professorial

consciousness is the kind of alter

alteration uh you get a much more varied

much

less predictable but much richer view of

what's going on