2 The Theory of how people figure out stuff from data humanoids seminar 1, fall ‘09 Tuesday, September 6, 2011 what it’s going to be about: what people (or other agents) need in order to make sense of the world but first: a little motivation why we should care: Tuesday, September 6, 2011 programming an intelligent humanoid is the same problem as programming an intelligent human (that is, like evolution, we have to figure out how to make sense of the world in order to act in it) The primary goal of this reading group to identify the research topics in cognitive science that help us make sense of how to make sense of the world. These goals relate strongly to our goals in humanoid robotics. Research in robotics shares a lot of fundamental ideas with various topics in psychology, particularly developmental work, and some of the most notable advances in the field have drawn heavily from concepts which originated in psychology. In psych there has been a huge amount of empirical and theoretical work on {what mechanisms underly human mental machinery, what the prereqs are for cognitive development, what the developmental arc is for acquiring these skills, even theories of how inference and decision are computed} and so i’d say that psych is more mature as a science than the fluffy descriptive field as labeled by its detractors. If science matures through an arc over descriptive, empirical, and theoretical phases, then what is robotics? Really, the 4th phase - application. In considering problems we want humanoid robots to solve, we're clearly a bit out of our depth on the theory. not surprisingly, many of the famous research domains in robotics were originally motivated by theories of mind: notably subsumption architecture from society of mind, and social robotics from theory of mind. why we should care: = Tuesday, September 6, 2011 programming an intelligent humanoid is the same problem as programming an intelligent human (that is, like evolution, we have to figure out how to make sense of the world in order to act in it) The primary goal of this reading group to identify the research topics in cognitive science that help us make sense of how to make sense of the world. These goals relate strongly to our goals in humanoid robotics. Research in robotics shares a lot of fundamental ideas with various topics in psychology, particularly developmental work, and some of the most notable advances in the field have drawn heavily from concepts which originated in psychology. In psych there has been a huge amount of empirical and theoretical work on {what mechanisms underly human mental machinery, what the prereqs are for cognitive development, what the developmental arc is for acquiring these skills, even theories of how inference and decision are computed} and so i’d say that psych is more mature as a science than the fluffy descriptive field as labeled by its detractors. If science matures through an arc over descriptive, empirical, and theoretical phases, then what is robotics? Really, the 4th phase - application. In considering problems we want humanoid robots to solve, we're clearly a bit out of our depth on the theory. not surprisingly, many of the famous research domains in robotics were originally motivated by theories of mind: notably subsumption architecture from society of mind, and social robotics from theory of mind. just a few examples of borrowed concepts... Tuesday, September 6, 2011 just a few examples of borrowed concepts... • social learning and theory of mind Tuesday, September 6, 2011 just a few examples of borrowed concepts... • social learning and theory of mind • the shape bias in object recognition Tuesday, September 6, 2011 just a few examples of borrowed concepts... • social learning and theory of mind • the shape bias in object recognition • object permanence Tuesday, September 6, 2011 just a few examples of borrowed concepts... • social learning and theory of mind • the shape bias in object recognition • object permanence • causal interaction Tuesday, September 6, 2011 just a few examples of borrowed concepts... • social learning and theory of mind • the shape bias in object recognition • object permanence • causal interaction • gaze detection Tuesday, September 6, 2011 just a few examples of borrowed concepts... • social learning and theory of mind • the shape bias in object recognition • object permanence • causal interaction • gaze detection • behavioral modularity Tuesday, September 6, 2011 and considering what we take for granted: we have a lot to learn from ourselves Tuesday, September 6, 2011 structure: (tree: taxonomy, continuum: political spectrum, cycles: seasons, relational: "better", "older", "taller" etc. (e.g. understanding of transitive property not understood till age 7) types: (e.g. distinguish horses from non-horses after just a few examples as a child) causality: (e.g. hands push blocks, not the other way around. ) and considering what we take for granted: • represent what other agents know, what they mean, and what they want we have a lot to learn from ourselves Tuesday, September 6, 2011 structure: (tree: taxonomy, continuum: political spectrum, cycles: seasons, relational: "better", "older", "taller" etc. (e.g. understanding of transitive property not understood till age 7) types: (e.g. distinguish horses from non-horses after just a few examples as a child) causality: (e.g. hands push blocks, not the other way around. ) and considering what we take for granted: • represent what other agents know, what they mean, and what they want • understand structured relationships between things we have a lot to learn from ourselves Tuesday, September 6, 2011 structure: (tree: taxonomy, continuum: political spectrum, cycles: seasons, relational: "better", "older", "taller" etc. (e.g. understanding of transitive property not understood till age 7) types: (e.g. distinguish horses from non-horses after just a few examples as a child) causality: (e.g. hands push blocks, not the other way around. ) and considering what we take for granted: • represent what other agents know, what they mean, and what they want • understand structured relationships between things • can learn to recognize novel types based on small, high-dimensional data we have a lot to learn from ourselves Tuesday, September 6, 2011 structure: (tree: taxonomy, continuum: political spectrum, cycles: seasons, relational: "better", "older", "taller" etc. (e.g. understanding of transitive property not understood till age 7) types: (e.g. distinguish horses from non-horses after just a few examples as a child) causality: (e.g. hands push blocks, not the other way around. ) and considering what we take for granted: • represent what other agents know, what they mean, and what they want • understand structured relationships between things • • can learn to recognize novel types based on small, high-dimensional data reason about causes and effects in the physical world we have a lot to learn from ourselves Tuesday, September 6, 2011 structure: (tree: taxonomy, continuum: political spectrum, cycles: seasons, relational: "better", "older", "taller" etc. (e.g. understanding of transitive property not understood till age 7) types: (e.g. distinguish horses from non-horses after just a few examples as a child) causality: (e.g. hands push blocks, not the other way around. ) and considering what we take for granted: • represent what other agents know, what they mean, and what they want • understand structured relationships between things • • can learn to recognize novel types based on small, high-dimensional data reason about causes and effects in the physical world • can use tools to extend your physical capabilities we have a lot to learn from ourselves Tuesday, September 6, 2011 structure: (tree: taxonomy, continuum: political spectrum, cycles: seasons, relational: "better", "older", "taller" etc. (e.g. understanding of transitive property not understood till age 7) types: (e.g. distinguish horses from non-horses after just a few examples as a child) causality: (e.g. hands push blocks, not the other way around. ) The Theory Theory Tuesday, September 6, 2011 read highlight on p2 now, her paper is an argument which you either read or you didn’t. I’m just going to summarize the theory The Theory Theory • An account of how we know Tuesday, September 6, 2011 read highlight on p2 now, her paper is an argument which you either read or you didn’t. I’m just going to summarize the theory The Theory Theory • An account of how we know • nativism vs. empiricism Tuesday, September 6, 2011 read highlight on p2 now, her paper is an argument which you either read or you didn’t. I’m just going to summarize the theory The Theory Theory • An account of how we know • nativism vs. empiricism • theory is empiricist 2 Tuesday, September 6, 2011 read highlight on p2 now, her paper is an argument which you either read or you didn’t. I’m just going to summarize the theory The Theory Theory • An account of how we know • nativism vs. empiricism • theory is empiricist • “an alternative to the innateness 2 hypothesis” Tuesday, September 6, 2011 read highlight on p2 now, her paper is an argument which you either read or you didn’t. I’m just going to summarize the theory The Theory Theory • An account of how we know • nativism vs. empiricism • theory is empiricist • “an alternative to the innateness 2 hypothesis” • Chomsky says: you can’t learn anything without knowing something Tuesday, September 6, 2011 read highlight on p2 now, her paper is an argument which you either read or you didn’t. I’m just going to summarize the theory The Innate Hypothesis Tuesday, September 6, 2011 2: can only formulate a small set of the possible representations and rules The Innate Hypothesis • One specific nativist hypothesis Tuesday, September 6, 2011 2: can only formulate a small set of the possible representations and rules The Innate Hypothesis • One specific nativist hypothesis • Proposed that human minds are highly constrained innately Tuesday, September 6, 2011 2: can only formulate a small set of the possible representations and rules The Innate Hypothesis • One specific nativist hypothesis • Proposed that human minds are highly constrained innately • Information from outside can trigger enrichment of representation, but not different constraints Tuesday, September 6, 2011 2: can only formulate a small set of the possible representations and rules Why such a radical theory? Tuesday, September 6, 2011 Why such a radical theory? • Poverty of the stimulus Tuesday, September 6, 2011 Why such a radical theory? • Poverty of the stimulus • Ancient philosophical argument: Tuesday, September 6, 2011 Why such a radical theory? • Poverty of the stimulus • Ancient philosophical argument: • all known learning mechanisms are too weak to explain human learning Tuesday, September 6, 2011 Why such a radical theory? • Poverty of the stimulus • Ancient philosophical argument: • all known learning mechanisms are too weak to explain human learning • Evidence for Chomsky came from language (proposed “universal grammar”) Tuesday, September 6, 2011 2 The Theory Tuesday, September 6, 2011 take 2 The Theory • Tuesday, September 6, 2011 take Fine, but how do you explain science (or other forms of human learning through data and experimentation)?? 2 The Theory take • Fine, but how do you explain science (or other forms of human learning through data and experimentation)?? • Instead: an empiricist alternative Tuesday, September 6, 2011 2 The Theory take • Fine, but how do you explain science (or other forms of human learning through data and experimentation)?? • • Instead: an empiricist alternative Tuesday, September 6, 2011 Gopnik says: 2 The Theory take • Fine, but how do you explain science (or other forms of human learning through data and experimentation)?? • • Instead: an empiricist alternative Tuesday, September 6, 2011 Gopnik says: • 1) humans DO have the ability to discover new systems of representation (theories) 2 The Theory take • Fine, but how do you explain science (or other forms of human learning through data and experimentation)?? • • Instead: an empiricist alternative Tuesday, September 6, 2011 Gopnik says: • 1) humans DO have the ability to discover new systems of representation (theories) • 2) these mechanisms explain human development What exactly is a “theory” Tuesday, September 6, 2011 ...in the way a perceptual system assigns representations to visual input or the syntactic system assigns representations to phonological input. read the “Neurath’s metaphor on sailboat”, p5 What exactly is a “theory” • “ a particular kind of system that assigns representations to inputs...” Tuesday, September 6, 2011 ...in the way a perceptual system assigns representations to visual input or the syntactic system assigns representations to phonological input. read the “Neurath’s metaphor on sailboat”, p5 What exactly is a “theory” • “ a particular kind of system that assigns representations to inputs...” • key difference w/ Chomskyan grammar: Tuesday, September 6, 2011 ...in the way a perceptual system assigns representations to visual input or the syntactic system assigns representations to phonological input. read the “Neurath’s metaphor on sailboat”, p5 What exactly is a “theory” • “ a particular kind of system that assigns representations to inputs...” • key difference w/ Chomskyan grammar: • the patterns of representation that occur can alter the nature of the representation system itself (yes neil, we know lisp does that) Tuesday, September 6, 2011 ...in the way a perceptual system assigns representations to visual input or the syntactic system assigns representations to phonological input. read the “Neurath’s metaphor on sailboat”, p5 Possible empiricist alternatives? Tuesday, September 6, 2011 Possible empiricist alternatives? • Behaviorism Tuesday, September 6, 2011 Possible empiricist alternatives? • Behaviorism • Championed most by Skinner (innateness was largely a reaction against this) Tuesday, September 6, 2011 Possible empiricist alternatives? • Behaviorism • Championed most by Skinner (innateness was largely a reaction against this) • Connectionism (McClelland & Rummelhart) Tuesday, September 6, 2011 Possible empiricist alternatives? • Behaviorism • Championed most by Skinner (innateness was largely a reaction against this) • Connectionism (McClelland & Rummelhart) • Modern alternative Tuesday, September 6, 2011 Possible empiricist alternatives? • Behaviorism • Championed most by Skinner (innateness was largely a reaction against this) • Connectionism (McClelland & Rummelhart) • Modern alternative • proposes neural mechanisms for abstraction Tuesday, September 6, 2011 What’s wrong with these? Tuesday, September 6, 2011 • Tuesday, September 6, 2011 What’s wrong with these? Both resort to a “scaling up” argument: • Tuesday, September 6, 2011 What’s wrong with these? Both resort to a “scaling up” argument: • Given some learning mechanism which can produce a representation that is at least somewhat more abstract or structured than the input, all that’s left is ‘scaling it up’ by enriching input or adding connections and *poof* we get human cognition • • Tuesday, September 6, 2011 What’s wrong with these? Both resort to a “scaling up” argument: • Given some learning mechanism which can produce a representation that is at least somewhat more abstract or structured than the input, all that’s left is ‘scaling it up’ by enriching input or adding connections and *poof* we get human cognition Problem: has never really worked • • Tuesday, September 6, 2011 What’s wrong with these? Both resort to a “scaling up” argument: • Given some learning mechanism which can produce a representation that is at least somewhat more abstract or structured than the input, all that’s left is ‘scaling it up’ by enriching input or adding connections and *poof* we get human cognition Problem: has never really worked • Output is never as complex or structured as the phenomena that we’re trying to explain So then how does it get started? Tuesday, September 6, 2011 So then how does it get started? • Starting-State Nativism (Gopnik) Tuesday, September 6, 2011 So then how does it get started? • Starting-State Nativism (Gopnik) • A set of innate constraints to guide theory formation Tuesday, September 6, 2011 So then how does it get started? • Starting-State Nativism (Gopnik) • A set of innate constraints to guide theory formation • But isn’t that like Chomsky? Tuesday, September 6, 2011 So then how does it get started? • Starting-State Nativism (Gopnik) • A set of innate constraints to guide theory formation • But isn’t that like Chomsky? • No, in this version the representation is subject to revision Tuesday, September 6, 2011 Thanks, but an example would be great... Tuesday, September 6, 2011 Thanks, but an example would be great... • Sorry, but there still isn’t a fully elaborated “theory making machine” Tuesday, September 6, 2011 Thanks, but an example would be great... • Sorry, but there still isn’t a fully elaborated “theory making machine” • But! There have been advances in learning theory (spurred by ML), that directly support the Theory2 Tuesday, September 6, 2011 Thanks, but an example would be great... • Sorry, but there still isn’t a fully elaborated “theory making machine” • But! There have been advances in learning theory (spurred by ML), that directly support the Theory2 • Tuesday, September 6, 2011 The skeleton of a theory Tuesday, September 6, 2011 * science is just a rational, institutionalized version of the theory2 The skeleton of a theory • Remember: the goal is to define a system that can incorporate new data in an analogous fashion that scientists do Tuesday, September 6, 2011 * science is just a rational, institutionalized version of the theory2 The skeleton of a theory • Remember: the goal is to define a system that can incorporate new data in an analogous fashion that scientists do • So, what do scientists do? Tuesday, September 6, 2011 * science is just a rational, institutionalized version of the theory2 The skeleton of a theory • Remember: the goal is to define a system that can incorporate new data in an analogous fashion that scientists do • So, what do scientists do? • Reason about cause and effect Tuesday, September 6, 2011 * science is just a rational, institutionalized version of the theory2 The skeleton of a theory • Remember: the goal is to define a system that can incorporate new data in an analogous fashion that scientists do • So, what do scientists do? • Reason about cause and effect • We need a system that builds representations of causal relationships from experience in the world Tuesday, September 6, 2011 * science is just a rational, institutionalized version of the theory2 The skeleton of a theory • Remember: the goal is to define a system that can incorporate new data in an analogous fashion that scientists do • So, what do scientists do? • Reason about cause and effect • We need a system that builds representations of causal relationships from experience in the world • note: this is really an introduction to the TICS paper, entitled “Mechanisms of theory formation in young children”, so keep that in mind as we go... Tuesday, September 6, 2011 * science is just a rational, institutionalized version of the theory2 Representing Cause and effect Tuesday, September 6, 2011 Representing Cause and effect • Core idea: capture it all in a bayes net Tuesday, September 6, 2011 Representing Cause and effect • Core idea: capture it all in a bayes net • Edges reflect conditional dependence (conditional independence implied by lack of edge) Tuesday, September 6, 2011 Mechanisms of theory formation in young children Tuesday, September 6, 2011 Learning about conditional probabilities Tuesday, September 6, 2011 Learning about conditional probabilities • Tuesday, September 6, 2011 Paradigm: which object caused the sound? Learning about conditional probabilities • Tuesday, September 6, 2011 Paradigm: which object caused the sound? • Children see various combinations of objects placed on a machine Learning about conditional probabilities • Tuesday, September 6, 2011 Paradigm: which object caused the sound? • Children see various combinations of objects placed on a machine • sometimes machine generates light & noise when block is placed on it Learning about conditional probabilities • Tuesday, September 6, 2011 Paradigm: which object caused the sound? • Children see various combinations of objects placed on a machine • sometimes machine generates light & noise when block is placed on it • Children are told “blickets make the machine go”, and then asked to point out the blicket Experiment 1: conditional independence asymmetry • P(E,B|A) = 2/3 • P(E,A|B) = 2/3 • P(E|A)P(B|A) = 2/3 • P(E|B)P(A|B) = 1/3 Tuesday, September 6, 2011 kids say A not B Experiment II: conditional independence symmetry P(E,B|A) = 0 P(E,A|B) = 0 P(E|A)P(B|A) = 0 P(E|B)P(A|B) = 0 Tuesday, September 6, 2011 kids equally likely to say A as B Experiment III: inference condition • P(E|B)=1 • P(E|A) = 1/2 • B fully explains activation of machine Tuesday, September 6, 2011 kids answer B Experiment IV: ambiguous condition • P(E|A) = 1 • P(E|B) = 1 Tuesday, September 6, 2011 Kids answer equally, or are unsure