Laura Schulz: The surprisingly logical minds of babies

Summary

In this talk, Laura Schulz discusses the remarkable ability of children to make inferences and learn from limited data, drawing parallels between scientific reasoning and early cognitive development. Through experiments involving babies, she demonstrates how infants can generalize information and engage in causal reasoning based on statistical evidence. Schulz emphasizes that human cognition, especially in children, is incredibly powerful and fundamentally different from machine learning. She argues for greater investment in nurturing children's cognitive development, as they represent humanity's greatest learning potential.

Takeaways

• 🧠 The ability to learn quickly from sparse, noisy data is a key trait shared by both scientists and children.
• 📊 Babies can generalize knowledge from small, representative samples, just like scientists do with data.
• 🎲 Babies consider whether their evidence is randomly sampled or cherry-picked, which affects their expectations about the world.
• 🔵 In an experiment, babies generalized squeaking behavior to different colored balls, but only when the sample seemed randomly drawn.
• 👶 Babies use limited statistical data to reason causally, deciding between fixing a problem themselves or asking for help.
• 🧩 Babies’ ability to learn from small samples extends to a wide range of things, such as causal relationships, objects, and even language.
• 🤖 The speaker contrasts human cognition with machine learning, highlighting how human minds make complex inferences from minimal data.
• 👨‍👩‍👧 The speaker emphasizes the importance of investing in children's development, given their extraordinary learning abilities.
• 💡 The speaker challenges the notion that human minds are inherently flawed, pointing out that humans make countless correct decisions effortlessly.
• 👨‍🏫 The structured, hierarchical nature of human knowledge is a key challenge that sets human cognition apart from current machine learning capabilities.

Q & A

What was the central problem of cognitive science highlighted by Mark Twain’s quote?

Mark Twain's quote humorously points out how science can make broad inferences from limited data. The central problem in cognitive science, as highlighted in the talk, is understanding how humans, especially children, can make accurate generalizations from small and noisy data.

How does the ability of children to generalize from sparse data relate to science?

Children, like scientists, can draw rich, abstract inferences from limited and noisy data. This ability is key in both fields: scientists use small samples to predict large-scale phenomena, and children use small experiences to form expectations about the world.

What was the first experiment discussed, and what did it demonstrate about babies' reasoning?

The first experiment involved showing babies two different statistical sampling conditions with blue and yellow balls. It demonstrated that babies are more likely to generalize properties (like squeaking) to other objects when the evidence appears randomly sampled, similar to how scientists assess evidence.

What does the result of the first experiment suggest about babies' understanding of probability?

The experiment suggests that babies, as young as 15 months, understand the concept of probability. They can infer that if evidence seems cherry-picked, it might not represent the larger population, leading them to limit their generalizations.

What is the ‘problem of confounded evidence’ discussed in the second experiment?

The problem of confounded evidence refers to the challenge of determining whether a failure (like a toy not working) is due to a flaw in the person using it or the object itself. This experiment tested whether babies could resolve this ambiguity using limited statistical data.

What were the key findings from the second experiment on causal reasoning?

The experiment showed that babies use small amounts of evidence to decide between two different strategies: either asking for help (changing the person) or trying a new tool (changing the toy). This demonstrates their ability to make causal inferences from minimal data.

How does this research challenge the narrative of human cognitive biases?

The speaker argues that, while humans are sometimes fallible and prone to biases, the bigger story is that human cognition is extraordinary. From a very young age, we make complex, accurate inferences and navigate the world successfully, which should be recognized as a remarkable achievement.

How does this research relate to the broader fields of neuroscience and machine learning?

Although neuroscience and machine learning have made significant strides, particularly with large amounts of data, this research emphasizes the unique ability of human minds to learn and infer from small amounts of data, something machines and even modern neuroscience have yet to fully replicate.

What makes the computational power of a human child different from current technologies?

Human children's minds are capable of making rich, structured inferences from minimal data. Unlike current technologies, which often rely on large datasets, human cognition involves hierarchical and structured knowledge that remains unmatched by machines.

What is the key takeaway from this research on children's learning for the future of innovation?

The key takeaway is that human children's learning abilities are unparalleled, and investing in their development—through education and caregiving—is as important as investing in technological innovation. By supporting their growth, we are planning for a better future.

Outlines

00:00

🧠 The Fascination of Science and Cognitive Inference

Mark Twain's witty remark on science highlights its ability to draw broad conclusions from minimal evidence, similar to how scientists deduce dinosaur existence from bones or cosmic compositions from spectral lines. This mirrors how children make abstract inferences from limited data. In the speaker's lab, they investigate how children rapidly learn complex concepts from sparse information, much like scientists generalize from small data samples. The research focuses on generalization and causal reasoning in child cognition.

05:00

👶 Babies' Ability to Generalize Evidence

The speaker introduces an experiment exploring how babies generalize information. Babies are shown a box of mostly blue balls, and they witness three blue balls being pulled out that squeak when squeezed. This evidence leads babies to expect that other balls, including yellow ones, might also squeak. This demonstrates that babies, much like scientists, make generalizations based on the representativeness of the data they observe.

10:02

🎥 Babies' Responses to Non-Random Sampling

In a follow-up experiment, babies observe the same blue balls being pulled from a box of mostly yellow balls. This non-random sampling suggests the blue balls were chosen deliberately, leading babies to believe only the blue balls squeak. The results show that babies, even at 15 months old, adjust their behavior based on the likelihood of evidence being representative, mirroring scientific reasoning.

15:05

🔄 Babies as Causal Reasoners

The speaker shifts to a problem of causal reasoning, showing how babies decide whether a toy is malfunctioning or if they are using it incorrectly. By observing others’ successes and failures with the toy, babies make informed decisions about whether to change the toy or the person using it. The experiment demonstrates that babies, from an early age, use small amounts of data to distinguish between causes and adapt their actions accordingly.

20:06

🧩 Learning from Small Data to Solve Big Problems

The speaker emphasizes that children’s ability to learn from limited data underlies much of human cultural knowledge. Babies use minimal examples to learn about tools, causal relationships, and even language. This cognitive ability forms the foundation of human ingenuity, allowing us to make sense of the world from small, sparse evidence.

🔬 Understanding Human Minds Through Science

In concluding, the speaker contrasts the current era’s focus on neuroscience, big data, and machine learning with the extraordinary abilities of the human mind. Human cognition, especially children’s capacity to learn and create from minimal information, represents a computational power that surpasses even the most advanced machines. Investing in nurturing young minds is essential for building a better future, as human minds have unique capabilities that technology may never replicate.

Keywords

💡Generalization

Generalization refers to the ability to extend a learned concept or behavior to new situations or examples. In the video, this concept is illustrated through babies learning to expect that certain balls will squeak after seeing a few examples of squeaky balls. This is directly tied to the video's theme of how children, like scientists, can draw broad conclusions from limited data.

💡Causal reasoning

Causal reasoning is the process of identifying cause-and-effect relationships. In the video, babies demonstrate causal reasoning by discerning whether a toy's failure to work is due to the toy itself or the person using it. This showcases how children, even from a young age, can use minimal evidence to form conclusions about the world.

💡Statistical evidence

Statistical evidence refers to data collected from observations or experiments that help to make predictions or generalizations. The video emphasizes how babies are sensitive to whether a sample of evidence (e.g., blue balls squeaking) is random or cherry-picked, indicating that even infants rely on statistical evidence to form their expectations.

💡Sparse data

Sparse data refers to having a limited amount of information from which to draw conclusions. The video focuses on how children can make accurate inferences from sparse, noisy data, such as learning from just a few examples or observations. This ability is contrasted with how machines often need vast amounts of data to achieve similar results.

💡Learning

Learning is the process of acquiring new knowledge or skills through experience, study, or teaching. The video explores how children, even at very young ages, are able to learn complex concepts like generalization and causality from minimal evidence. This highlights the powerful learning mechanisms present in human cognition.

💡Random sampling

Random sampling is the process of selecting items or data points in such a way that every item has an equal chance of being chosen. The video discusses how babies are more likely to generalize information when they believe the evidence they see (such as three squeaky blue balls) is drawn randomly, suggesting they understand the importance of representative sampling.

💡Structured knowledge

Structured knowledge refers to organized, hierarchical systems of information. In the video, this concept is brought up to emphasize how human minds, especially children's, use structured knowledge to make sense of the world. This allows them to learn rapidly from minimal examples, unlike machines, which often rely on large data sets.

💡Imitation

Imitation is the process of learning by observing and replicating the behavior of others. In the video, babies imitate the researcher by squeezing balls that squeak, showing how children use imitation as a learning tool to understand their environment and form hypotheses about how things work.

💡Machine learning

Machine learning is a field of artificial intelligence that focuses on enabling machines to learn from data and improve over time without explicit programming. The video contrasts the efficiency of children's learning with that of machine learning systems, noting that while machines need large amounts of data, children can learn from much smaller sets.

💡Cognitive science

Cognitive science is the interdisciplinary study of the mind and its processes, including how people think, learn, and remember. The video, presented by a cognitive scientist, delves into how children develop cognitive skills such as generalization and causal reasoning, providing insight into the human capacity for rapid, abstract learning from limited data.

Highlights

Mark Twain humorously captures a key problem in cognitive science: the ability to infer much from little data, highlighting the essence of scientific inquiry.

Children, like scientists, make rich inferences from sparse, noisy data—inferring abstract concepts quickly and accurately from minimal evidence.

The speaker discusses research on how children generalize knowledge, such as learning properties of objects (e.g., ducks float, balls bounce) from limited examples.

In experiments, babies generalize properties (such as squeaking) from a few observations if they believe the sample is representative of a broader population.

When the evidence appears cherry-picked (e.g., selecting only blue balls from a mostly yellow set), babies behave differently, suggesting they intuitively understand biased sampling.

Babies can distinguish between plausibly random and deliberate sampling and use this distinction to form different expectations about the world.

Even with one blue ball from a mostly yellow box, babies assume that squeaking might apply to all items, showing how they form hypotheses from limited evidence.

Babies use minimal data to infer causal relationships, such as determining whether failure to activate a toy is due to the person or the object itself.

In one experiment, babies change toys if they observe equal failure from multiple people, indicating they can infer that the problem lies with the toy rather than the user.

In another scenario, babies change people rather than toys if the data suggests that certain individuals succeed while others fail with the same object.

By their second year, babies make strategic decisions about when to explore new objects or ask for help based on the statistical evidence they observe.

The research demonstrates that babies, like scientists, develop expectations and causal reasoning from sparse evidence—using their observations to guide their interactions with the world.

The speaker emphasizes that while neuroscience and big data are advancing, humans, especially children, have an innate ability to learn from small amounts of information.

The ability of human minds to think of new ideas, make discoveries, and generate art and literature is highlighted as one of humanity's most profound abilities.

Investing in the development of children's cognitive abilities, through education and care, is compared to investing in other forms of technology for a better future.

Transcripts

Mark Twain summed up what I take to be

one of the fundamental problems of cognitive science

with a single witticism.

He said, "There's something fascinating about science.

One gets such wholesale returns of conjecture

out of such a trifling investment in fact."

(Laughter)

Twain meant it as a joke, of course, but he's right:

There's something fascinating about science.

From a few bones, we infer the existence of dinosuars.

From spectral lines, the composition of nebulae.

From fruit flies,

the mechanisms of heredity,

and from reconstructed images of blood flowing through the brain,

or in my case, from the behavior of very young children,

we try to say something about the fundamental mechanisms

of human cognition.

In particular, in my lab in the Department of Brain and Cognitive Sciences at MIT,

I have spent the past decade trying to understand the mystery

of how children learn so much from so little so quickly.

Because, it turns out that the fascinating thing about science

is also a fascinating thing about children,

which, to put a gentler spin on Mark Twain,

is precisely their ability to draw rich, abstract inferences

rapidly and accurately from sparse, noisy data.

I'm going to give you just two examples today.

One is about a problem of generalization,

and the other is about a problem of causal reasoning.

And although I'm going to talk about work in my lab,

this work is inspired by and indebted to a field.

I'm grateful to mentors, colleagues, and collaborators around the world.

Let me start with the problem of generalization.

Generalizing from small samples of data is the bread and butter of science.

We poll a tiny fraction of the electorate

and we predict the outcome of national elections.

We see how a handful of patients responds to treatment in a clinical trial,

and we bring drugs to a national market.

But this only works if our sample is randomly drawn from the population.

If our sample is cherry-picked in some way --

say, we poll only urban voters,

or say, in our clinical trials for treatments for heart disease,

we include only men --

the results may not generalize to the broader population.

So scientists care whether evidence is randomly sampled or not,

but what does that have to do with babies?

Well, babies have to generalize from small samples of data all the time.

They see a few rubber ducks and learn that they float,

or a few balls and learn that they bounce.

And they develop expectations about ducks and balls

that they're going to extend to rubber ducks and balls

for the rest of their lives.

And the kinds of generalizations babies have to make about ducks and balls

they have to make about almost everything:

shoes and ships and sealing wax and cabbages and kings.

So do babies care whether the tiny bit of evidence they see

is plausibly representative of a larger population?

Let's find out.

I'm going to show you two movies,

one from each of two conditions of an experiment,

and because you're going to see just two movies,

you're going to see just two babies,

and any two babies differ from each other in innumerable ways.

But these babies, of course, here stand in for groups of babies,

and the differences you're going to see

represent average group differences in babies' behavior across conditions.

In each movie, you're going to see a baby doing maybe

just exactly what you might expect a baby to do,

and we can hardly make babies more magical than they already are.

But to my mind the magical thing,

and what I want you to pay attention to,

is the contrast between these two conditions,

because the only thing that differs between these two movies

is the statistical evidence the babies are going to observe.

We're going to show babies a box of blue and yellow balls,

and my then-graduate student, now colleague at Stanford, Hyowon Gweon,

is going to pull three blue balls in a row out of this box,

and when she pulls those balls out, she's going to squeeze them,

and the balls are going to squeak.

And if you're a baby, that's like a TED Talk.

It doesn't get better than that.

(Laughter)

But the important point is it's really easy to pull three blue balls in a row

out of a box of mostly blue balls.

You could do that with your eyes closed.

It's plausibly a random sample from this population.

And if you can reach into a box at random and pull out things that squeak,

then maybe everything in the box squeaks.

So maybe babies should expect those yellow balls to squeak as well.

Now, those yellow balls have funny sticks on the end,

so babies could do other things with them if they wanted to.

They could pound them or whack them.

But let's see what the baby does.

(Video) Hyowon Gweon: See this? (Ball squeaks)

Did you see that? (Ball squeaks)

Cool.

See this one?

(Ball squeaks)

Wow.

Laura Schulz: Told you. (Laughs)

(Video) HG: See this one? (Ball squeaks)

Hey Clara, this one's for you. You can go ahead and play.

(Laughter)

LS: I don't even have to talk, right?

All right, it's nice that babies will generalize properties

of blue balls to yellow balls,

and it's impressive that babies can learn from imitating us,

but we've known those things about babies for a very long time.

The really interesting question

is what happens when we show babies exactly the same thing,

and we can ensure it's exactly the same because we have a secret compartment

and we actually pull the balls from there,

but this time, all we change is the apparent population

from which that evidence was drawn.

This time, we're going to show babies three blue balls

pulled out of a box of mostly yellow balls,

and guess what?

You [probably won't] randomly draw three blue balls in a row

out of a box of mostly yellow balls.

That is not plausibly randomly sampled evidence.

That evidence suggests that maybe Hyowon was deliberately sampling the blue balls.

Maybe there's something special about the blue balls.

Maybe only the blue balls squeak.

Let's see what the baby does.

(Video) HG: See this? (Ball squeaks)

See this toy? (Ball squeaks)

Oh, that was cool. See? (Ball squeaks)

Now this one's for you to play. You can go ahead and play.

(Fussing) (Laughter)

LS: So you just saw two 15-month-old babies

do entirely different things

based only on the probability of the sample they observed.

Let me show you the experimental results.

On the vertical axis, you'll see the percentage of babies

who squeezed the ball in each condition,

and as you'll see, babies are much more likely to generalize the evidence

when it's plausibly representative of the population

than when the evidence is clearly cherry-picked.

And this leads to a fun prediction:

Suppose you pulled just one blue ball out of the mostly yellow box.

You [probably won't] pull three blue balls in a row at random out of a yellow box,

but you could randomly sample just one blue ball.

That's not an improbable sample.

And if you could reach into a box at random

and pull out something that squeaks, maybe everything in the box squeaks.

So even though babies are going to see much less evidence for squeaking,

and have many fewer actions to imitate

in this one ball condition than in the condition you just saw,

we predicted that babies themselves would squeeze more,

and that's exactly what we found.

So 15-month-old babies, in this respect, like scientists,

care whether evidence is randomly sampled or not,

and they use this to develop expectations about the world:

what squeaks and what doesn't,

what to explore and what to ignore.

Let me show you another example now,

this time about a problem of causal reasoning.

And it starts with a problem of confounded evidence

that all of us have,

which is that we are part of the world.

And this might not seem like a problem to you, but like most problems,

it's only a problem when things go wrong.

Take this baby, for instance.

Things are going wrong for him.

He would like to make this toy go, and he can't.

I'll show you a few-second clip.

And there's two possibilities, broadly:

Maybe he's doing something wrong,

or maybe there's something wrong with the toy.

So in this next experiment,

we're going to give babies just a tiny bit of statistical data

supporting one hypothesis over the other,

and we're going to see if babies can use that to make different decisions

about what to do.

Here's the setup.

Hyowon is going to try to make the toy go and succeed.

I am then going to try twice and fail both times,

and then Hyowon is going to try again and succeed,

and this roughly sums up my relationship to my graduate students

in technology across the board.

But the important point here is it provides a little bit of evidence

that the problem isn't with the toy, it's with the person.

Some people can make this toy go,

and some can't.

Now, when the baby gets the toy, he's going to have a choice.

His mom is right there,

so he can go ahead and hand off the toy and change the person,

but there's also going to be another toy at the end of that cloth,

and he can pull the cloth towards him and change the toy.

So let's see what the baby does.

(Video) HG: Two, three. Go! (Music)

LS: One, two, three, go!

Arthur, I'm going to try again. One, two, three, go!

YG: Arthur, let me try again, okay?

One, two, three, go! (Music)

Look at that. Remember these toys?

See these toys? Yeah, I'm going to put this one over here,

and I'm going to give this one to you.

You can go ahead and play.

LS: Okay, Laura, but of course, babies love their mommies.

Of course babies give toys to their mommies

when they can't make them work.

So again, the really important question is what happens when we change

the statistical data ever so slightly.

This time, babies are going to see the toy work and fail in exactly the same order,

but we're changing the distribution of evidence.

This time, Hyowon is going to succeed once and fail once, and so am I.

And this suggests it doesn't matter who tries this toy, the toy is broken.

It doesn't work all the time.

Again, the baby's going to have a choice.

Her mom is right next to her, so she can change the person,

and there's going to be another toy at the end of the cloth.

Let's watch what she does.

(Video) HG: Two, three, go! (Music)

Let me try one more time. One, two, three, go!

Hmm.

LS: Let me try, Clara.

One, two, three, go!

Hmm, let me try again.

One, two, three, go! (Music)

HG: I'm going to put this one over here,

and I'm going to give this one to you.

You can go ahead and play.

(Applause)

LS: Let me show you the experimental results.

On the vertical axis, you'll see the distribution

of children's choices in each condition,

and you'll see that the distribution of the choices children make

depends on the evidence they observe.

So in the second year of life,

babies can use a tiny bit of statistical data

to decide between two fundamentally different strategies

for acting in the world:

asking for help and exploring.

I've just shown you two laboratory experiments

out of literally hundreds in the field that make similar points,

because the really critical point

is that children's ability to make rich inferences from sparse data

underlies all the species-specific cultural learning that we do.

Children learn about new tools from just a few examples.

They learn new causal relationships from just a few examples.

They even learn new words, in this case in American Sign Language.

I want to close with just two points.

If you've been following my world, the field of brain and cognitive sciences,

for the past few years,

three big ideas will have come to your attention.

The first is that this is the era of the brain.

And indeed, there have been staggering discoveries in neuroscience:

localizing functionally specialized regions of cortex,

turning mouse brains transparent,

activating neurons with light.

A second big idea

is that this is the era of big data and machine learning,

and machine learning promises to revolutionize our understanding

of everything from social networks to epidemiology.

And maybe, as it tackles problems of scene understanding

and natural language processing,

to tell us something about human cognition.

And the final big idea you'll have heard

is that maybe it's a good idea we're going to know so much about brains

and have so much access to big data,

because left to our own devices,

humans are fallible, we take shortcuts,

we err, we make mistakes,

we're biased, and in innumerable ways,

we get the world wrong.

I think these are all important stories,

and they have a lot to tell us about what it means to be human,

but I want you to note that today I told you a very different story.

It's a story about minds and not brains,

and in particular, it's a story about the kinds of computations

that uniquely human minds can perform,

which involve rich, structured knowledge and the ability to learn

from small amounts of data, the evidence of just a few examples.

And fundamentally, it's a story about how starting as very small children

and continuing out all the way to the greatest accomplishments

of our culture,

we get the world right.

Folks, human minds do not only learn from small amounts of data.

Human minds think of altogether new ideas.

Human minds generate research and discovery,

and human minds generate art and literature and poetry and theater,

and human minds take care of other humans:

our old, our young, our sick.

We even heal them.

In the years to come, we're going to see technological innovations

beyond anything I can even envision,

but we are very unlikely

to see anything even approximating the computational power of a human child

in my lifetime or in yours.

If we invest in these most powerful learners and their development,

in babies and children

and mothers and fathers

and caregivers and teachers

the ways we invest in our other most powerful and elegant forms

of technology, engineering and design,

we will not just be dreaming of a better future,

we will be planning for one.

Thank you very much.

(Applause)

Chris Anderson: Laura, thank you. I do actually have a question for you.

First of all, the research is insane.

I mean, who would design an experiment like that? (Laughter)

I've seen that a couple of times,

and I still don't honestly believe that that can truly be happening,

but other people have done similar experiments; it checks out.

The babies really are that genius.

LS: You know, they look really impressive in our experiments,

but think about what they look like in real life, right?

It starts out as a baby.

Eighteen months later, it's talking to you,

and babies' first words aren't just things like balls and ducks,

they're things like "all gone," which refer to disappearance,

or "uh-oh," which refer to unintentional actions.

It has to be that powerful.

It has to be much more powerful than anything I showed you.

They're figuring out the entire world.

A four-year-old can talk to you about almost anything.

(Applause)

CA: And if I understand you right, the other key point you're making is,

we've been through these years where there's all this talk

of how quirky and buggy our minds are,

that behavioral economics and the whole theories behind that

that we're not rational agents.

You're really saying that the bigger story is how extraordinary,

and there really is genius there that is underappreciated.

LS: One of my favorite quotes in psychology

comes from the social psychologist Solomon Asch,

and he said the fundamental task of psychology is to remove

the veil of self-evidence from things.

There are orders of magnitude more decisions you make every day

that get the world right.

You know about objects and their properties.

You know them when they're occluded. You know them in the dark.

You can walk through rooms.

You can figure out what other people are thinking. You can talk to them.

You can navigate space. You know about numbers.

You know causal relationships. You know about moral reasoning.

You do this effortlessly, so we don't see it,

but that is how we get the world right, and it's a remarkable

and very difficult-to-understand accomplishment.

CA: I suspect there are people in the audience who have

this view of accelerating technological power

who might dispute your statement that never in our lifetimes

will a computer do what a three-year-old child can do,

but what's clear is that in any scenario,

our machines have so much to learn from our toddlers.

LS: I think so. You'll have some machine learning folks up here.

I mean, you should never bet against babies or chimpanzees

or technology as a matter of practice,

but it's not just a difference in quantity,

it's a difference in kind.

We have incredibly powerful computers,

and they do do amazingly sophisticated things,

often with very big amounts of data.

Human minds do, I think, something quite different,

and I think it's the structured, hierarchical nature of human knowledge

that remains a real challenge.

CA: Laura Schulz, wonderful food for thought. Thank you so much.

LS: Thank you. (Applause)