Building Machines That Learn and Think Like People

Building Machines That Learn and Think Like People Brenden M. Lake Tomer D. Ullman Joshua B. Tenenbaum Samuel J. Gershman

Tag layer of the DHSS Meta-Tagging Project. The article text is reproduced from the staged copy named in sourceDesc and is not altered; the tags are held separately in standOff so that overlapping and nested spans are preserved.

Building Machines That Learn and Think Like People 2017 Behavioral and Brain Sciences / arXiv 10.1017/S0140525X16001837 1604.00289 lake2017-machines-learn-think lake2017-machines-learn-think.txt, the project's staged copy 47434ed8b778f38af9bff014487d5590bb1ba834 arxiv-open cognitive-psychology

the paper's own words

the tag layer; a sentence written by the tagger about the paper, always anchored to a quotation but not itself a quotation

0 spans located by searching the stored text for the quote's exact words, allowing any whitespace between them critique Quotation withheld because this record is not cleared for redistribution. human-like learning and thinking in machines Quotation withheld because this record is not cleared for redistribution. review review Quotation withheld because this record is not cleared for redistribution. mechanisms PERFORMANCE PARITY IS GRANTED AND SIMILARITY DENIED IN CONSECUTIVE SENTENCES: these systems achieve performance that EQUALS OR EVEN BEATS HUMANS IN SOME RESPECTS - DESPITE their biological inspiration and performance achievements, they DIFFER FROM HUMAN INTELLIGENCE IN CRUCIAL WAYS. Matching a score is separated from being alike, and the concession comes first Quotation withheld because this record is not cleared for redistribution. mechanisms THE CRITERION IS SPLIT INTO TWO DIMENSIONS SO THAT A SYSTEM CAN FAIL ON EITHER: truly HUMAN-LIKE learning and thinking machines will have to reach beyond current engineering trends in BOTH WHAT THEY LEARN, AND HOW THEY LEARN IT. One adjective, two independent requirements Quotation withheld because this record is not cleared for redistribution. mechanisms A QUESTION IS DECLARED OLDER THAN THE NAMES OF BOTH FIELDS THAT ASK IT: the questions of whether and how AI should relate to human cognitive psychology are OLDER THAN THE TERMS ARTIFICIAL INTELLIGENCE AND COGNITIVE PSYCHOLOGY. The problem existed before either vocabulary did, which means neither vocabulary was built to state it Quotation withheld because this record is not cleared for redistribution. mechanisms THE POSITION BEING ARGUED AGAINST IS QUOTED FROM A TEXTBOOK, AS AN ANALOGY ABOUT ABANDONING IMITATION: the quest for ARTIFICIAL FLIGHT succeeded WHEN THE WRIGHT BROTHERS AND OTHERS STOPPED IMITATING BIRDS and started using wind tunnels and learning about aerodynamics. The rival view is presented in its own strongest form, in its own words, with a page reference Quotation withheld because this record is not cleared for redistribution. mechanisms A DEFENDER'S HOPE IS GRANTED SYMPATHY AND THEN DOUBTED, WITH THE DOUBT MARKED AS UNCERTAIN: Quotation withheld Disagreement stated as an uncertainty about the opponent's prospects rather than as a refutation Quotation withheld because this record is not cleared for redistribution. mechanisms THE BENCHMARK'S SCALE IS MADE CONCRETE BY NAMING THE OBJECTS: benchmark tasks provide HUNDREDS OR THOUSANDS OF EXAMPLES PER CLASS - 1000 HAIRBRUSHES, 1000 PINEAPPLES, etc. - and one dataset includes 6000 examples of each handwritten digit. An abstract quantity is turned into a picture of what the quantity is made of Quotation withheld because this record is not cleared for redistribution. mechanisms TWO QUANTITIES A SINGLE COMPARISON CONFLATES ARE SEPARATED IN A FOOTNOTE: the time required to train the network - **COMPUTE TIME** - IS NOT THE SAME AS THE GAME (EXPERIENCE) TIME, and compute time CAN BE LONGER. Two clocks under one word, distinguished where a reader might not look Quotation withheld because this record is not cleared for redistribution. mechanisms A METRIC IS DESCRIBED AS DESIGNED TO BLOCK ONE SPECIFIC CHEAT, AND ITS RIVALS' STATUS IS DECLARED UNKNOWN: the reported scores use the HUMAN STARTS measure, DESIGNED TO PREVENT NETWORKS FROM JUST MEMORIZING LONG SEQUENCES OF SUCCESSFUL ACTIONS FROM A SINGLE STARTING POINT - and faster learning and higher scores HAVE BEEN REPORTED USING OTHER METRICS, BUT IT IS UNCLEAR HOW WELL THE NETWORKS ARE GENERALIZING WITH THESE ALTERNATIVE METRICS. The number depends on which measure, and only one of them is known to mean what it says Quotation withheld because this record is not cleared for redistribution. mechanisms AN OPPOSING SCHOOL'S POSITION IS SUMMARISED AS A CLAIM ABOUT METAPHOR: proponents maintain that classic types of structured knowledge - graphs, grammars, rules, objects, programs - CAN BE USEFUL YET MISLEADING METAPHORS for characterizing thought, and that these structures are MORE EPIPHENOMENAL THAN REAL. The disagreement is over whether the vocabulary names things or describes them Quotation withheld because this record is not cleared for redistribution. mechanisms THE OBJECTION TO THE PAPER'S OWN COMPARISON IS RAISED BY THE PAPER, EARLY, AND FLAGGED AS ANTICIPATED: one might object that both challenges draw an UNFAIR COMPARISON between the speed of human learning and neural network learning - discussed in detail later, BUT WE FEEL IT IS IMPORTANT TO ANTICIPATE HERE AS WELL. The strongest attack on the paper's method is placed before the method is used Quotation withheld because this record is not cleared for redistribution. mechanisms AND THE REFRAMING THE PAPER ADOPTS IS QUOTED FROM AN ANONYMOUS REVIEWER OF AN EARLIER DRAFT: THEY MAY BE BETTER SEEN AS SOLVING DIFFERENT TASKS - the human ENCOUNTERING ONE IN A YEARS-LONG STRING OF PROBLEMS WITH RICH OVERLAPPING STRUCTURE, the network STARTING COMPLETELY FROM SCRATCH. The comparison fails not through a confound but because one phrase, `the task`, was covering two different things - and the authors reply WE AGREE about this quote: EXTENDED so the span carries the authors' assent in full - "We agree, and indeed this is another way of putting our point here" - rather than stopping at "We agree". The two words alone would have left it unclear whether they were conceding the objection or adopting the reframing. They adopt it: an anonymous reviewer's sentence becomes the paper's own statement of its point. NOTE ALSO that the reviewer quotation has a PAGE NUMBER spliced inside it a few words earlier - "solving the same task 14 differently" - which is why the span starts at the following sentence. Quotation withheld because this record is not cleared for redistribution. mechanisms A CAPACITY IS SHOWN TO PERSIST DESPITE THE PRACTICE THAT WOULD EXPLAIN IT: children are FAR MORE PRACTICED than adults at learning new concepts - ROUGHLY NINE OR TEN NEW WORDS EACH DAY - YET the ability for rapid ONE-SHOT learning DOES NOT DISAPPEAR IN ADULTHOOD. The obvious explanation for the skill is present in one group and the skill in both Quotation withheld because this record is not cleared for redistribution. mechanisms THE PAPER'S OWN CRITERIA ARE CALLED INGREDIENTS AND PLACED AFTER A REVIEW OF OTHER PEOPLE'S: first the criteria PREVIOUSLY OFFERED by cognitive scientists, developmental psychologists and AI researchers are reviewed - THEN what the authors VIEW AS THE ESSENTIAL INGREDIENTS is articulated. The list is presented as a synthesis with a named provenance, not as a definition Quotation withheld because this record is not cleared for redistribution. planning in complex (model-based) systems is cumbersome and slow for real-time control what this limits: the real-time control claim: PLANNING IN COMPLEX MODELS IS CUMBERSOME AND SLOW Quotation withheld because this record is not cleared for redistribution. The article says very little about natural language, a capacity the authors concede is central to human intelligence. what this limits: the scope of the whole article: WE HAVE SAID LITTLE ... ABOUT PEOPLE'S ABILITY TO COMMUNICATE AND THINK IN NATURAL LANGUAGE, A DISTINCTIVELY HUMAN COGNITIVE CAPACITY WHERE MACHINE CAPABILITIES LAG STRIKINGLY Quotation withheld because this record is not cleared for redistribution. The proposed list of core ingredients is explicitly not claimed to be complete. what this limits: the list of core ingredients: WE DID NOT INTEND OUR LIST TO BE COMPLETE Quotation withheld because this record is not cleared for redistribution. The authors concede the objection that humans bring far richer prior experience to a new task than a network does, which complicates the learning-speed comparisons the paper rests on. what this limits: the human-machine comparison: humans bring far richer prior experience to a new task than a network does, which the authors concede as an objection Quotation withheld because this record is not cleared for redistribution. Their own prior model of the Characters Challenge does not capture the full human ability. Quotation withheld because this record is not cleared for redistribution. Equal performance on the benchmark is stated not to indicate equal capability. what this limits: any comparison of human and network ability based on these benchmarks: the authors state that performing equally well on MNIST and other large-scale image classification tasks does not mean humans and networks learn and think in the same way Quotation withheld because this record is not cleared for redistribution. The human-level comparison is stated for specific data sets, with one described as within reach rather than reached. what this limits: any claim that networks have matched human performance: the authors report an error rate very close to human-level on one task and describe human-level performance as within reach on the ImageNet benchmark Quotation withheld because this record is not cleared for redistribution. the simulation engine that renders a game world is used as the model of intuitive physics: the mind is proposed to run something like it to predict what happens next in a scene Quotation withheld because this record is not cleared for redistribution. the representations and algorithms of the connectionist approach were drawn from neuroscience rather than from cognitive psychology, and the paper marks that as the choice which shaped it Quotation withheld because this record is not cleared for redistribution. MNIST MNIST Quotation withheld because this record is not cleared for redistribution. richer representations richer representations Quotation withheld because this record is not cleared for redistribution. human-level performance human-level performance Quotation withheld because this record is not cleared for redistribution. ImageNet ImageNet Quotation withheld because this record is not cleared for redistribution. neural networks neural networks Quotation withheld because this record is not cleared for redistribution. people people Quotation withheld because this record is not cleared for redistribution. deep learning deep learning Quotation withheld because this record is not cleared for redistribution. PDP models PDP models Quotation withheld because this record is not cleared for redistribution. Intuitive psychology Intuitive psychology Quotation withheld because this record is not cleared for redistribution. equal benchmark performance - NOT EQUAL CAPABILITY: the authors state that humans and networks may score the same on MNIST and other large-scale image classification tasks while not learning or thinking in the same way, and name two respects in which they differ - the number of examples needed and the richness of what is learned Quotation withheld because this record is not cleared for redistribution. human-level performance - A PHRASE ATTACHED TO PARTICULAR DATA SETS: reported as an error rate very close to human on one task and as within reach on another, so the expression names a relation to a specific benchmark rather than a general parity Quotation withheld because this record is not cleared for redistribution. the list of differences - A FLOOR, NOT A SET: the authors write that there are AT LEAST two important differences before naming two, so the enumeration is explicitly open and a reader who treats the two as the complete account has closed a list the authors left open Quotation withheld because this record is not cleared for redistribution. `deep learning` / `PDP models` - A RENAMING ACROSS THIRTY YEARS, AND THE PAPER SAYS THE CONTENT DID NOT CHANGE WITH THE NAME. `The resurgence of recent interest in neural networks, more commonly referred to as “deep learning,” share the same representational commitments and often even the same learning algorithms as the earlier PDP models.` Both names describe the same objects and neither is wrong; what changed is which property the field chose to put in the name - PDP named the computation (parallel, distributed), `deep` names the layer count. The paper then glosses the new name explicitly: `“Deep” refers to the fact that more powerful models can be built by composing many layers of representation.` Quotation withheld because this record is not cleared for redistribution. contradicts → the view that current deep-learning systems capture human intelligence Quotation withheld because this record is not cleared for redistribution. builds-on → Lake, Salakhutdinov, & Tenenbaum (2015) probabilistic program induction for handwritten characters Quotation withheld because this record is not cleared for redistribution. contradicts → V. Mnih et al. (2015) Deep Q-Network playing Atari games Quotation withheld because this record is not cleared for redistribution. contradicts → AlphaGo (Silver et al., 2016) as a system that understands the game Quotation withheld because this record is not cleared for redistribution. neural networks —contrasted-with→ people neural networks people differs in: how many examples are needed and how rich the learned representation is - a list the authors mark as at least two · strength of the relation: 1 Quotation withheld because this record is not cleared for redistribution. deep learning —renamed-across-fields→ PDP models deep learning PDP models the relation itself is separately evidenced · strength of the relation: 0.9 · note: Both names sit inside the one span, as the type requires. Weight 0.9 rather than 1.0: the paper says 'often even the same learning algorithms' - the identity is asserted for the commitments and hedged for the algorithms. The split is across TIME rather than across a field boundary, which is the nearest type available and is recorded here as such. · the relation stated: neural networks, more commonly referred to as “deep learning,” share the same representational commitments and often even the same learning algorithms as the earlier PDP models Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but the span says deep learning shares the representational commitments of the earlier PDP models; renamed-across-fields is loose. Quotation withheld because this record is not cleared for redistribution. 3 — proposed ingredients (a,b,c) for human-like machines Quotation withheld because this record is not cleared for redistribution. ~half — Krizhevsky convnet cut the prior error rate Quotation withheld because this record is not cleared for redistribution. late 1980s — when HMMs became the leading speech approach Quotation withheld because this record is not cleared for redistribution. these systems differ in crucial ways from how humans think Quotation withheld because this record is not cleared for redistribution. human-like machines should build causal models of the world Quotation withheld because this record is not cleared for redistribution. learning should be grounded in intuitive physics and psychology Quotation withheld because this record is not cleared for redistribution. machines should harness compositionality and learning-to-learn Quotation withheld because this record is not cleared for redistribution. causal models causal models Quotation withheld because this record is not cleared for redistribution. intuitive theories intuitive theories Quotation withheld because this record is not cleared for redistribution. compositionality compositionality Quotation withheld because this record is not cleared for redistribution. deep learning deep learning Quotation withheld because this record is not cleared for redistribution. deep reinforcement learning deep reinforcement learning Quotation withheld because this record is not cleared for redistribution. convolutional neural network convolutional neural network Quotation withheld because this record is not cleared for redistribution. learning-to-learn learning-to-learn Quotation withheld because this record is not cleared for redistribution. Bayesian Bayesian Quotation withheld because this record is not cleared for redistribution. mean mean Quotation withheld because this record is not cleared for redistribution. EVIDENCE CORRECTED (RE-GROUNDED) · regression regression EVIDENCE CORRECTED: RE-GROUNDED. The old evidence was 'Quotation withheld' - THE TITLE OF A PAPER BY SOMEBODY ELSE, sitting in this paper's bibliography. An internal tool cuts the reference list on purpose, for exactly this reason, so the row was invisible to every grounding check. lake2017 uses 'regression' once in its own prose and that sentence is now the evidence. Quotation withheld because this record is not cleared for redistribution. statistical-significance statistical-significance Quotation withheld because this record is not cleared for redistribution.

The full text of this article is not redistributable, so it is not included here. The evidence sentences are withheld with it; the tag layer and its grounding status remain. The article itself is at https://doi.org/10.1017/S0140525X16001837.