DreamCoder: growing generalizable, interpretable knowledge with wake–sleep Bayesian program learning

Philosophical Transactions of the Royal Society A Mathematical Physical and Engineering Sciences, Journal Year: 2023, Volume and Issue: 381(2251)

Published: June 4, 2023

Expert problem-solving is driven by powerful languages for thinking about problems and their solutions. Acquiring expertise means learning these languages—systems of concepts, alongside the skills to use them. We present DreamCoder, a system that learns solve writing programs. It builds creating domain-specific programming expressing domain together with neural networks guide search programs within languages. A ‘wake–sleep’ algorithm alternately extends language new symbolic abstractions trains network on imagined replayed problems. DreamCoder solves both classic inductive tasks creative such as drawing pictures building scenes. rediscovers basics modern functional programming, vector algebra classical physics, including Newton’s Coulomb’s laws. Concepts are built compositionally from those learned earlier, yielding multilayered representations interpretable transferrable tasks, while still growing scalably flexibly experience. This article part discussion meeting issue ‘Cognitive artificial intelligence’.

Language: Английский

---

Concepts at the Interface

Oxford University Press eBooks, Journal Year: 2024, Volume and Issue: unknown

Published: Aug. 9, 2024

Abstract Research on concepts has concentrated the way people apply online, when presented with a stimulus. Just as important, however, is use of offline, planning what to do or thinking about case. There strong evidence that inferences driven by conceptual thought draw heavily special-purpose resources: sensory, motoric, affective, and evaluative. At same time, afford general-purpose recombination support domain-general reasoning processes—phenomena have long been focus philosophers. growing consensus theory must encompass both kinds process. This book shows how are able act an interface between systems. Concept-driven can take advantage complementary costs benefits each. The lays out empirically-based account different ways in which takes us new conclusions underpins planning, decision-making, action. It also spells three useful implications account. First, it allows reconstruct commonplace idea draws meaning concept. Second, offers insight into human cognition avoids frame problem complementary, less discussed, ‘if-then problem’ for nested processing dispositions. Third, metacognition concept-driven various ways. framework developed elucidates makes especially powerful cognitive resource.

Language: Английский

---

Perceiving Topological Relations

Psychological Science, Journal Year: 2025, Volume and Issue: unknown

Published: Feb. 18, 2025

There are many ways to describe and represent the visuospatial world. A space can be described by its euclidean properties—the size of objects, angles boundaries, distances between them. also in nonspatial terms: One could explain layout a city order streets. Somewhere between, topological representations—such as those commonly depicted public-transit maps—capture coarse relational structure without precise detail, offering relatively efficient, low-dimensional way capturing spatial content. Here, we ask whether human adults quickly automatically perceive such relations. In six experiments, show that differences simple features influence range visual tasks from object matching number estimation search. We discuss possibility relations kind primitive supports representation.

Language: Английский

---

Humans parsimoniously represent auditory sequences by pruning and completing the underlying network structure

eLife, Journal Year: 2023, Volume and Issue: 12

Published: May 2, 2023

Successive auditory inputs are rarely independent, their relationships ranging from local transitions between elements to hierarchical and nested representations. In many situations, humans retrieve these dependencies even limited datasets. However, this learning at multiple scale levels is poorly understood. Here, we used the formalism proposed by network science study representation of higher-order structures interaction in sequences. We show that human adults exhibited biases perception elements, which made them sensitive high-order such as communities. This behavior consistent with creation a parsimonious simplified model evidence they receive, achieved pruning completing elements. observation suggests brain does not rely on exact memories but world. Moreover, bias can be analytically modeled memory/efficiency trade-off. correctly accounts for previous findings, including transition probabilities well structures, unifying sequence across scales. finally propose putative implementations bias.

Language: Английский

---

Brain-imaging evidence for compression of binary sound sequences in human memory

eLife, Journal Year: 2023, Volume and Issue: 12

Published: Nov. 1, 2023

According to the language-of-thought hypothesis, regular sequences are compressed in human memory using recursive loops akin a mental program that predicts future items. We tested this theory by probing for 16-item made of two sounds. recorded brain activity with functional MRI and magneto-encephalography (MEG) while participants listened hierarchy variable complexity, whose minimal description required transition probabilities, chunking, or nested structures. Occasional deviant sounds probed participants’ knowledge sequence. predicted task difficulty would be proportional complexity derived from length our formal language. Furthermore, should increase learned sequences, decrease deviants. These predictions were upheld both fMRI MEG, indicating sequence highly dependent on structure become weaker delayed as increases. The proposed language recruited bilateral superior temporal, precentral, anterior intraparietal, cerebellar cortices. regions overlapped extensively localizer mathematical calculation, much less spoken written processing. propose these areas collectively encode repetitions variations their composition into

Language: Английский

---

Models of Possibilities Instead of Logic as the Basis of Human Reasoning

Minds and Machines, Journal Year: 2024, Volume and Issue: 34(3)

Published: June 4, 2024

Language: Английский

---

Universal conceptual modeling: principles, benefits, and an agenda for conceptual modeling research

Software & Systems Modeling, Journal Year: 2024, Volume and Issue: 23(5), P. 1077 - 1100

Published: Sept. 3, 2024

Abstract The paper proposes universal conceptual modeling , that strives to be as general-purpose possible and accessible anyone, professionals non-experts alike. idea of is meant catalyze new thinking in used evaluate develop solutions, such languages, approaches for requirements elicitation, or tools. These solutions should usable by many people design agents purposes possible, aspiring the ideals modeling. We propose foundations form six principles: flexibility, accessibility, ubiquity, minimalism, primitivism, modularity. then demonstrate utility these principles existing languages understand practices. Finally, we future research opportunities realize

Language: Английский

---

Discovering Symbolic Cognitive Models from Human and Animal Behavior

bioRxiv (Cold Spring Harbor Laboratory), Journal Year: 2025, Volume and Issue: unknown

Published: Feb. 6, 2025

Symbolic models play a key role in cognitive science, expressing computationally precise hypotheses about how the brain implements process. Identifying an appropriate model typically requires great deal of effort and ingenuity on part human scientist. Here, we adapt FunSearch Romera-Paredes et al. (2024), recently developed tool that uses Large Language Models (LLMs) evolutionary algorithm, to automatically discover symbolic accurately capture animal behavior. We consider datasets from three species performing classic reward-learning task has been focus substantial modeling effort, find discovered programs outperform state-of-the-art for each. The can readily be interpreted as cognition, instantiating interpretable learning decision-making algorithms. Broadly, these results demonstrate viability using LLM-powered program synthesis propose novel scientific regarding mechanisms cognition.

Language: Английский

---

Children’s Understanding of Topological Relations

Open Mind, Journal Year: 2025, Volume and Issue: 9, P. 401 - 417

Published: Jan. 1, 2025

A core aim of developmental cognitive science is to uncover the basic building blocks human thought. For instance, work revealing that even young children, adults without formal education, and distant animal species are sensitive Euclidean properties indicates humans may be endowed with some primitive understanding geometry. But what about other forms geometry? Here, we explore children's sensitivity topological spatial forms. We show like adults, spontaneously distinguish match items in accordance their relations. As well, judgments object similarity remarkably consistent adults', indicating stability concepts throughout lifespan. Finally, compare various geometric curvature, perpendicularity, symmetry, find while there variability performance across all features tested, overall for vs. comparable. Collectively, these findings suggest children have an intuitive relations among visuospatial representation.

Language: Английский

---

Crows recognize geometric regularity

Science Advances, Journal Year: 2025, Volume and Issue: 11(15)

Published: April 11, 2025

The perception of geometric regularity in shapes, a form elementary Euclidean geometry, is fundamental mathematical intuition humans. We demonstrate this understanding an animal, the carrion crow. Crows were trained to detect visually distinct intruder shape among six concurrent arbitrary shapes. crows able immediately apply concept quadrilaterals, identifying one that exhibited differing properties compared others set. effect, showing better performance with shapes featuring right angles, parallel lines, or symmetry over more irregular This advantage did not require learning. Our findings suggest intuitions are specific humans but deeply rooted biological evolution.

Language: Английский

---