PaperlayerView exact revision on arXiv

Source-linked AI summary

Building Machines That Learn and Think Like People

Brenden M. Lake, Tomer D. Ullman, Joshua B. Tenenbaum, Samuel J. Gershman

arXiv:1604.00289v3cs.AIcs.CVcs.LGcs.NEstat.ML

TL;DR

Recent deep-learning systems achieve strong results across recognition and control, yet differ from people in their learning and representational capacities. The paper synthesizes cognitive science into a program centered on causal models, intuitive theories, compositionality, and learning-to-learn, while proposing hybrid neural and structured approaches. Its examples show strong benchmark performance alongside substantial experience requirements and scope limitations.

  • Problem

    High-performing deep-learning systems still differ from human intelligence in how they learn and represent knowledge.

    Source: §1 · p. 2, §1.1 · p. 3, §3.2 · pp. 13–14

  • Method

    The paper synthesizes cognitive-science findings and proposes integrating neural networks with causal models, intuitive theories, compositionality, learning-to-learn, and efficient inference.

    Source: §1 · p. 2, §1.2 · p. 6, §4.1.1 · p. 18

  • Results

    Deep-learning systems achieve strong results, including 0.2% error on MNIST and 96% of a professional gamer’s score, while requiring much more experience than humans.

    Source: §3.1 · p. 9, §3.2 · p. 12

  • Takeaways & Limitations

    More human-like AI should combine neural-network strengths with structured cognitive models and richer inductive biases.

    Source: §1 · p. 2, §1.2 · p. 6

    The authors are unsure that larger datasets, richer tasks, and more computation alone can recover the cognitive ingredients supplied by human evolution and learning.

    Source: §5 · p. 35

Abstract

from arXiv · show

Recent progress in artificial intelligence (AI) has renewed interest in building systems that learn and think like people. Many advances have come from using deep neural networks trained end-to-end in tasks such as object recognition, video games, and board games, achieving performance that equals or even beats humans in some respects. Despite their biological inspiration and performance achievements, these systems differ from human intelligence in crucial ways. We review progress in cognitive science suggesting that 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. Specifically, we argue that these machines should (a) build causal models of the world that support explanation and understanding, rather than merely solving pattern recognition problems; (b) ground learning in intuitive theories of physics and psychology, to support and enrich the knowledge that is learned; and (c) harness compositionality and learning-to-learn to rapidly acquire and generalize knowledge to new tasks and situations. We suggest concrete challenges and promising routes towards these goals that can combine the strengths of recent neural network advances with more structured cognitive models.

1 Introduction

The paper examines what human-like learning and thinking require beyond recent deep-learning successes, arguing for causal models, intuitive theories, compositionality, and learning-to-learn. It proposes combining these cognitive ingredients with neural networks and model-free methods for efficient real-time inference and control.

1 Introduction

  • Deep learning has produced major gains in recognition, speech, and control, but these achievements do not by themselves establish human-like cognition.The paper reviews performance advances while distinguishing engineering success from emulating human learning and thought.Source: §1 · pp. 1–2
  • Human intelligence is characterized as model-building for explanation, understanding, imagination, and planning, rather than prediction alone in a specific task.Pattern recognition remains useful as a computational support for model-based cognition through model-free algorithms.Source: §1 · p. 2

1.1 What this article is not

  • The paper focuses on human-inspired AI and argues that reverse-engineering cognition remains useful because people outperform machines in concept learning, scene understanding, language, and general-purpose reasoning.The scope excludes AI approaches that do not seek cognitive or neural inspiration, while acknowledging neural networks’ important contributions.Source: §1.1 · pp. 3, 5

1.2 Overview of the key ideas

  • The proposed ingredients include developmental intuitive physics and psychology, causal model-building, compositionality, and learning-to-learn for efficient generalization from sparse experience.Early intuitive theories constrain predictions, while compositionality and learning-to-learn support rapid acquisition of richly structured models.Source: §1.2 · pp. 5–6
  • Human-like systems should combine rich model-building with neural networks that learn to perform inference and with model-based and model-free reinforcement learning.Model-based methods support planning but can be slow, whereas model-free methods enable fast control but may be inflexible or less accurate.Source: §1.2 · p. 6

2 Cognitive and neural inspiration in artificial intelligence

The paper situates contemporary deep learning within longer debates about whether cognition is best modeled through structured symbolic representations or flexible sub-symbolic neural computation. It argues that neural networks can emulate structured behavior, but human-like AI still depends on the nature, compositionality, transferability, and developmental origins of learned representations.

Cognitive and neural inspiration in artificial intelligence

  • Turing’s child-machine proposal anticipated learning through rewards and punishments, echoing behaviorism and the empiricist assumptions of modern connectionist models.This historical view treated much knowledge as learnable from statistical patterns of sensory inputs.Source: §2 · p. 6
  • Early AI drew heavily on cognitive psychology and symbolic representations for reasoning, language, planning, and vision, while parallel work developed neuron-like sub-symbolic computation.The General Problem Solver exemplified an attempt to build an AI algorithm that also modeled human problem solving.Source: §2 · pp. 6–7
  • The paper distinguishes pattern recognition from model building within the PDP tradition and focuses on learned representations’ form, compositionality, transferability, and developmental inductive biases.Recent large-scale discriminative systems more purely exemplify pattern recognition, whereas some earlier PDP work was closer to model building.Source: §2 · pp. 7–8
  • Generic neural networks have reproduced structured behaviors across language, physics, taxonomy, object recognition, and video games, but the paper asks whether this suffices for truly human-like learning and thinking.These systems can emulate rule-like behavior without necessarily implementing the structured representations that cognition requires.Source: §2 · p. 8

3 Challenges for building more human-like machines

The paper uses the Characters and Frostbite challenges to show that human-like intelligence requires learning rich, flexible concepts from little experience rather than merely matching patterns. These challenges expose differences in representation, planning, transfer, and learning efficiency between people and contemporary neural networks.

3.1 The Characters Challenge

  • MNIST systems reach near-human digit-recognition accuracy with 6,000 training images per class, whereas people learn new characters from far fewer examples.The best reported deep convolutional networks achieve a 0.2% error rate on MNIST.Source: §3.1 · p. 9
  • People can learn a novel handwritten character from one example and use its concept to classify, generate, parse, and compose related visual concepts.These abilities go beyond recognizing examples and reflect a richer representation of the underlying concept.Source: §3.1 · p. 9, Figure 1 · p. 10
  • Probabilistic program induction advances the Characters Challenge, but people still represent richer relations among strokes and integrate variants such as optional cross-bars into coherent concepts.The paper suggests combining deep learning with probabilistic program induction for richer versions of the challenge.Source: §3.1 · pp. 9–10

3.2 The Frostbite Challenge

  • Frostbite requires temporally extended planning because the agent must use moving, state-dependent ice floes to construct an igloo before time expires.The game also includes rewards and hazards that vary across levels.Source: §3.2 · pp. 10–11, Figure 2 · p. 11
  • The original DQN trained on 924 hours of unique game experience achieved less than 10% of human-level performance, while DQN+ and DQN++ reached 83% and 96%.The variants improved through smarter replay and, for DQN++, more efficient parameter sharing.Source: §3.2 · p. 12, Figure 3 · p. 13
  • Humans grasp Frostbite within minutes, whereas even improved DQNs require multiple hours to perform reliably above random play.The authors associate this contrast with differences in the kinds of knowledge or learning mechanisms used.Source: §3.2 · p. 12
  • Incremental rewards provide DQN with sub-goals, while people can infer the higher-level igloo-building goal without equivalent feedback.Sparse feedback is also identified as a difficulty for DQNs in games such as Montezuma’s Revenge.Source: §3.2 · p. 13
  • People can adapt learned game models to arbitrary new goals and explain the game, whereas neural networks generally require substantial retraining or reconfiguration for new tasks.The comparison highlights flexibility and transfer as central components of human intelligence.Source: §3.2 · pp. 13–15

4 Core ingredients of human intelligence

The paper identifies several core ingredients as key building blocks for human-like learning and thought, while emphasizing that they need not be innate. These ingredients may be built in, learned, or enriched through experience, but contemporary systems rarely capture them together.

4 Core ingredients of human intelligence

  • Integrating the proposed core ingredients could produce more powerful and more human-like learning and thinking abilities than current AI systems.The paper presents these ingredients as key building blocks rather than an exhaustive account of intelligence.Source: §4 · p. 15
  • A core ingredient may be innate, learned, or enriched through experience; the claim is that it actively contributes to human-like learning and thought.The paper remains agnostic about the origins of these ingredients.Source: §4 · p. 15
  • Human learners approach new characters and games with extensive real-world experience that contemporary deep learning systems generally lack.This prior experience is presented as difficult to emulate in a general sense.Source: §4 · p. 15

4.1 Developmental start-up software

The paper describes early-emerging intuitive theories of core domains, especially physics and psychology, as structured knowledge that supports flexible human learning and reasoning. It considers how these capacities might augment deep learning through causal, simulation-based representations.

4.1 Developmental start-up software

  • Early human cognition includes intuitive theories of number, space, physics, and psychology, organized around entities and abstract principles with causal structure.Children also seek informative data, isolate variables, test causal hypotheses, and learn selectively from others.Source: §4.1 · pp. 15–16

4.1.1 Intuitive physics

  • Intuitive physics represents objects, physical properties, and forces approximately and probabilistically, enabling mental simulations of immediate motion and flexible counterfactual predictions.Physics-engine models predict tower stability and fit adults’ quantitative judgments, while supporting varied interventions such as adding, removing, or altering blocks.Source: §4.1.1 · pp. 16–17, Figure 4 · p. 17
  • PhysNet matched human performance on simple real tower images and exceeded it on synthetic images, but required 100,000–200,000 scenes for one narrow task.Its limited generalization contrasts with people’s ability to handle novel judgments and complex scenes without task-specific retraining.Source: §4.1.1 · p. 18
  • Integrating intuitive physics with deep learning could reduce training data, improve performance, and avoid retraining when game objects’ behavior, rewards, or appearance change.Such representations could also ease adding new object types, such as bears in later Frostbite levels.Source: §4.1.1 · pp. 18–19

4.1.2 Intuitive psychology

  • The computational architecture of intuitive psychology remains disputed: cue-based accounts become cumbersome as scenarios change, whereas generative inverse-planning models represent goals, beliefs, efficiency, and agents.Simulation-based reasoning can flexibly handle novel goals and recursively model agents thinking about other agents.Source: §4.1.2 · pp. 19–20
  • A full account of intuitive psychology may require genuine representations of agency, goals, efficiency, and reciprocal relations, whose emergence from purely predictive deep networks remains unclear.Changing an agent’s setting, goals, or beliefs is difficult to generalize without understanding these concepts.Source: §4.1.2 · pp. 20–21
  • Intuitive psychology supports efficient social learning by inferring agents’ beliefs, desires, and intentions from observed behavior, such as learning that birds are dangerous without direct encounters.Explicit agent representations can also predict how a helpful sidekick will assist under new circumstances.Source: §4.1.2 · p. 21

4.2 Learning as rapid model building

This section presents human learning as rapid model building, emphasizing compositional, hierarchical, and causal representations that support prediction, action, explanation, imagination, composition, and transfer across tasks. It examines how structured prior knowledge and learning-to-learn enable rapid generalization, while noting that current systems remain slower than humans.

4.2 Learning as rapid model building

  • People learn most cognitively natural concepts from sparse data, whereas deep neural networks are data hungry and use information less efficiently on benchmarks such as ImageNet, MNIST, and Atari.Children can infer new concept boundaries from a few examples, while cited benchmarks provide hundreds or thousands of examples per class or hundreds of hours of gameplay.Source: §4.2 · p. 22
  • BPL achieved human-level one-shot handwritten-character classification, outperformed convolutional networks, and generated, parsed, and recombined novel concepts in human-like ways.Its stochastic programs represent causal, compositional procedures that generate concept examples, while probabilistic inference supports noise handling and creative tasks.Source: §4.2 · pp. 23–24

4.2.1 Compositionality

  • Compositional representations combine reusable primitives, parts, and relations to construct infinitely many concepts and support efficient learning and generalization to new tasks and configurations.For Frostbite, explicitly representing repeated objects and their combinations can improve economy and generalization across levels with different object arrangements.Source: §4.2.1 · pp. 25–26

4.2.2 Causality

  • Causal models represent how observations are produced, enriching perception and generalization; BPL’s motor-program representation supports one-shot classification and generation, unlike noncausal generative neural networks.Human-level scene understanding integrates intuitive physics, intuitive psychology, compositionality, and causal relationships, which deep-network captions often fail to capture.Source: §4.2.2 · pp. 27–28

4.2.3 Learning-to-learn

  • Learning-to-learn transfers priors and inductive biases across related tasks, but current neural networks require more pre-training than people or BPL and remain far from solving the Characters Challenge.BPL transfers reusable primitives, generative pieces, and learned variability across hierarchical levels, while some of its helpful structure remains built into the prior rather than learned.Source: §4.2.3 · pp. 29–30
  • For novel concepts such as a two-wheeled vehicle, people recombine learned parts and relations into new models, suggesting benefits from compositional, hierarchical, causal neural representations.Transferable sub-concepts include wheels, motors, handlebars, and relations such as attached and powered by.Source: §4.2.3 · p. 30
  • People transfer knowledge across perception, object types, causal relations, goals, and strategies, using prior hypotheses about game entities and physical constraints to guide exploration.This multilevel transfer supports generalization across new scenarios and can help people learn from a single mistake or avoid mistakes altogether.Source: §4.2.3 · pp. 30–31
  • BPL achieves human-level one-shot classification and generation after pre-training on only five alphabets, while deep classifiers retain 23% versus humans’ 4% error.With six times more data from 30 alphabets, neural-network error remains 30%, or two to three times higher than humans.Source: §4.2.3 · p. 30
  • Actor-mimic transfers knowledge from 13 Atari games, reducing new-game learning from four or five million frames to one or two million, yet humans reach comparable scores within minutes.The source describes roughly four million frames, or 18.5 hours, of expert-game experience per pre-training game.Source: §4.2.3 · p. 31
  • Rapid machine learning may require structured causal models grounded in intuitive physics and psychology, rather than deep systems trained on many games alone.Such models could transfer knowledge more efficiently and support faster learning of new games.Source: §4.2.3 · p. 31

4.3 Thinking Fast

Human-like intelligence requires reconciling rich structured models with fast inference and action. The section considers approximate inference and cooperation between model-based and model-free reinforcement learning as routes toward that goal.

4.3.1 Approximate inference in structured models

  • Rich causal and theory-like models support human-like reasoning but make inference computationally difficult, often requiring intractable search over program spaces.The challenge is especially severe when hypotheses are numerous and only a few fit the data.Source: §4.3.1 · pp. 31–32
  • People may guide hypothesis selection using abstract structural properties of problems, while Monte Carlo methods explain some cognitive phenomena but remain difficult to scale to program and theory learning.Children have been reported to use distributional and dynamical properties to constrain plausible hypotheses.Source: §4.3.1 · pp. 32–33
  • Amortized inference offers a route to efficiency by learning feed-forward mappings from questions to plausible answers, potentially integrating neural networks with probabilistic models and programs.Proposed approaches include recognition networks, variational optimization, nearest-neighbor density estimation, and differentiable programming.Source: §4.3.1 · p. 33

4.3.2 Model-based and model-free reinforcement learning

  • Human flexibility is attributed to model-based planning, which can recombine an environment model with new reward functions, whereas routine practice may shift control toward faster model-free processes.This trade-off can be understood as arbitration between flexibility and speed.Source: §4.3.2 · pp. 33–34
  • Efficiency may arise by distilling model-based plans into cached model-free values, allowing cooperation between flexible planning and fast action selection.The proposed process can occur offline and resembles consolidation in reinforcement learning.Source: §4.3.2 · p. 34
  • Intrinsic motivation may supplement external rewards by driving uncertainty reduction and construction of environmental models, although deep reinforcement learning is only beginning to address it.Internal value can depend on the agent’s current goal and mental state.Source: §4.3.2 · p. 34

5 Responses to common questions

The authors address critiques about biological plausibility, prior human experience, and whether larger-scale deep learning alone could yield human-like intelligence. They argue that cognitive ingredients and suitable representational structure remain central, while neuroscience currently provides limited constraints.

Prior experience

  • Human learning advantages reflect richer starting knowledge, which the authors treat as fundamental building blocks for rapid learning from sparse data.They remain agnostic about whether these ingredients arise mainly from genetic programming or experience-driven development.Source: §5 · p. 35
  • Learning-to-learn may support transfer, but human-like generalization requires compositional representations that capture causal structure and intuitive physics and psychology.Simply training conventional neural networks across related tasks may not suffice for novel-task generalization.Source: §5 · p. 35

Scaling deep learning

  • The authors argue that building human-like representations from scratch through weight-space gradient updates may be infeasible, motivating architectural structure search or built-in infant-like knowledge and inductive biases.They present structure-search methods and richer starting representations as promising alternatives, while acknowledging uncertainty about their practicality.Source: §5 · pp. 35–36
  • Any human-like AI approach could benefit from incorporating core cognitive ingredients, including additional structure or learning mechanisms beyond current deep-learning systems.The claim is intended to be orthogonal to whether the eventual system is based on deep learning.Source: §5 · p. 36

Biological plausibility

  • Biological plausibility should not determine cognitive ingredients by itself because neuroscience offers valuable inspiration but many neural-computation claims remain uncertain.The authors favor cognitive plausibility as a surer current foundation while expecting neuroscience to impose stronger constraints over time.Source: §5 · pp. 36–37
  • Backpropagation lacks an established brain implementation, but existing neuroscience neither definitively supports nor rules it out, and it has proven highly valuable for engineering.The authors therefore reject biological implausibility as sufficient reason to abandon the algorithm.Source: §5 · p. 36
  • Hebbian mechanisms face a stronger cognitive challenge because their timing and persistence properties do not match important behavioral learning patterns.The authors especially question whether purely Hebbian mechanisms could implement the cognitive ingredients proposed in the paper.Source: §5 · p. 37

3. Language is essential for human intelligence. Why is it not more prominent here?

The paper treats language as essential to fully human-like intelligence but focuses first on cognitive ingredients that support linguistic meaning and acquisition. It argues that language both builds on and extends intuitive reasoning, compositionality, and learning-to-learn.

Language as a future ingredient

  • The paper acknowledges that current neural networks remain far from human language abilities and asks how machines can acquire richer linguistic capacity.Language is identified as a distinctively human capacity that receives limited treatment in the article.Source: §3 · p. 37
  • The authors argue that intuitive physics, intuitive psychology, causal models, compositionality, and rapid learning precede language and provide building blocks for linguistic meaning and acquisition.They cite developmental and cognitive theories linking these capacities to language learning.Source: §3 · p. 37

Open questions

  • The paper does not claim its ingredient list is complete, leaving recursion, symbol reuse, and shared intentionality as possible additions for future research.These possibilities are presented as open questions rather than settled requirements.Source: §3 · pp. 37–38

Language as an amplifier

  • Language expands reasoning beyond infants’ immediate situations and strengthens learning-to-learn and compositionality by relating new concepts to existing ones.The authors conclude that language must be central to the full project of building human-like learning and thinking machines.Source: §3 · p. 38

6 Looking forward

The authors argue that human-like AI should be judged by learning efficiency, flexible generalization, causal understanding, and transfer beyond narrow task performance. They highlight structured neural systems, model-based search, and cognitive ingredients as routes toward that goal, while identifying major gaps between current systems and human capabilities.

6 Looking forward

  • People learn from less data and generalize more flexibly than current algorithms, which the authors attribute to causal and compositional representations.For handwritten characters, people can recognize, generate, and extend concepts from one or a few examples, unlike current deep networks.Source: §6 · p. 38
  • Deep learning may move toward human-like learning and thought by incorporating psychological ingredients such as those proposed in this paper.The authors frame this as a direction for future research rather than an achieved capability.Source: §6 · pp. 38–39

6.1 Promising directions in deep learning

  • Selective attention, augmented working memory, and experience replay illustrate a trend toward integrating cognitive insights with deep neural networks.These mechanisms are described as lower-level than the paper’s central ingredients but as promising foundations for higher-level integration.Source: §6.1 · p. 39
  • Attention improves processing by focusing models on selected input regions and coordinating complex sequential outputs, with gains reported in translation, recognition, and captioning.The authors suggest it could help build more causal generative neural models.Source: §6.1 · p. 39
  • Differentiable programming augments neural networks with external memory and executable structures, enabling tasks such as copying, sorting, graph navigation, and algorithm learning.These systems suggest a possible synthesis of program induction and gradient-based learning, though their learned representations remain relatively low-level.Source: §6.1 · pp. 39–40
  • AlphaGo demonstrated that deep convolutional networks and Monte Carlo Tree Search can be integrated to defeat a world-class Go player, but this achievement opens questions about human-like understanding and generalization.The paper treats the combination of pattern recognition and model-based search as an important accomplishment and research direction.Source: §6.1 · p. 40, Figure 7 · p. 41
  • AlphaGo trained on vastly more games than a human champion, motivating systems that could reach professional performance from roughly 50,000 games rather than tens of millions.The authors propose combining AlphaGo’s advances with complementary cognitive ingredients as one possible route.Source: §6.1 · p. 40
  • Human-like Go AI should learn quickly from instruction, observation, and a few games, then adapt flexibly across board sizes, geometries, rules, and player configurations.Current AlphaGo policies and value functions appear unlikely to generalize autonomously across such variants because they do not explicitly represent the game’s goals, rules, and constraints.Source: §6.1 · pp. 40, 42
Loading 1604.00289v3…