Spatial Relation Categorization in Infants and Deep Neural Networks DOI Open Access
Guy Davidson, A. Emin Orhan, Brenden M. Lake

et al.

Published: April 24, 2023

Spatial relations, such as above, below, between, and containment, are important mediators in children’s understanding of the world (Piaget, 1954). The development these relational categories infancy has been extensively studied (Quinn, 2003) yet little is known about their computational underpinnings. Using developmental tests, we examine extent to which deep neural networks, pretrained on a standard vision benchmark or egocentric video captured from one baby’s perspective, form categorical representations for visual stimuli depicting relations. Notably, networks did not receive any explicit training We then analyze whether recover similar patterns ones identified development, reproducing relative difficulty categorizing different spatial relations stimulus abstractions. find that evaluate tend many observed with simpler “above versus below” “between outside”, but struggle match findings related “containment”. identify factors choice model architecture, pretraining data, experimental design contribute patterns, highlight predictions made by our modeling results. Our results open door infants’ earliest categorization abilities modern machine learning tools demonstrate utility productivity this approach.

Language: Английский

Grounded language acquisition through the eyes and ears of a single child DOI
Wai Keen Vong, Wentao Wang, A. Emin Orhan

et al.

Science, Journal Year: 2024, Volume and Issue: 383(6682), P. 504 - 511

Published: Feb. 1, 2024

Starting around 6 to 9 months of age, children begin acquiring their first words, linking spoken words visual counterparts. How much this knowledge is learnable from sensory input with relatively generic learning mechanisms, and how requires stronger inductive biases? Using longitudinal head-mounted camera recordings one child aged 25 months, we trained a neural network on 61 hours correlated visual-linguistic data streams, feature-based representations cross-modal associations. Our model acquires many word-referent mappings present in the child’s everyday experience, enables zero-shot generalization new referents, aligns its linguistic conceptual systems. These results show critical aspects grounded word meaning are through joint representation associative input.

Language: Английский

Citations

41
A stimulus-computable rational model of habituation in infants and adults DOI Open Access
Gal Raz, Anjie Cao,

Rebecca Saxe

et al.

Published: Jan. 8, 2025

How do we decide what to look at and when stop looking? Even very young infants engage in active visual selection, looking less as stimuli are repeated (habituation) regaining interest novel subsequently introduced (dishabituation). The mechanisms underlying these time changes remain uncertain, however, due limits on both the scope of existing formal models empirical precision measurements infant behavior. To address this, developed Rational Action, Noisy Choice for Habituation (RANCH) model, which operates over raw images makes quantitative predictions participants’ behaviors. In a series pre-registered experiments, exposed adults varying durations measured familiar stimuli. We found that data were well captured by RANCH. Using RANCH’s stimulus-computability, also tested its out-of-sample about magnitude dishabituation new experiment manipulated similarity between stimulus. By framing behaviors rational decision-making, this work identified how dynamics learning exploration guide our attention from infancy through adulthood.

Language: Английский

Citations

0
A stimulus-computable rational model of habituation in infants and adults DOI Open Access
Gal Raz, Anjie Cao,

Rebecca Saxe

et al.

Published: Jan. 8, 2025

How do we decide what to look at and when stop looking? Even very young infants engage in active visual selection, looking less as stimuli are repeated (habituation) regaining interest novel subsequently introduced (dishabituation). The mechanisms underlying these time changes remain uncertain, however, due limits on both the scope of existing formal models empirical precision measurements infant behavior. To address this, developed Rational Action, Noisy Choice for Habituation (RANCH) model, which operates over raw images makes quantitative predictions participants’ behaviors. In a series pre-registered experiments, exposed adults varying durations measured familiar stimuli. We found that data were well captured by RANCH. Using RANCH’s stimulus-computability, also tested its out-of-sample about magnitude dishabituation new experiment manipulated similarity between stimulus. By framing behaviors rational decision-making, this work identified how dynamics learning exploration guide our attention from infancy through adulthood.

Language: Английский

Citations

0
Artificial intelligence tackles the nature–nurture debate DOI
Justin N. Wood

Nature Machine Intelligence, Journal Year: 2024, Volume and Issue: 6(4), P. 381 - 382

Published: April 19, 2024

Language: Английский

Citations

2
The Limitations of Large Language Models for Understanding Human Language and Cognition DOI Creative Commons
Christine Cuskley, Rebecca Woods,

Molly Flaherty

et al.

Open Mind, Journal Year: 2024, Volume and Issue: 8, P. 1058 - 1083

Published: Jan. 1, 2024

Researchers have recently argued that the capabilities of Large Language Models (LLMs) can provide new insights into longstanding debates about role learning and/or innateness in development and evolution human language. Here, we argue on two grounds LLMs alone tell us very little language cognition terms acquisition evolution. First, any similarities between output are purely functional. Borrowing "four questions" framework from ethology,

Language: Английский

Citations

2
Parallel development of social behavior in biological and artificial fish DOI Creative Commons
J. T. McGraw, Donsuk Lee, Justin N. Wood

et al.

Nature Communications, Journal Year: 2024, Volume and Issue: 15(1)

Published: Dec. 5, 2024

Abstract Our algorithmic understanding of vision has been revolutionized by a reverse engineering paradigm that involves building artificial systems perform the same tasks as biological systems. Here, we extend this to social behavior. We embodied neural networks in fish and raised virtual tanks mimicked rearing conditions fish. When had deep reinforcement learning curiosity-derived rewards, they spontaneously developed fish-like behaviors, including collective behavior preferences (favoring in-group over out-group members). The also naturalistic ocean worlds, showing these models generalize real-world contexts. Thus, animal-like behaviors can develop from generic algorithms (reinforcement intrinsic motivation). study provides foundation for reverse-engineering development using image-computable intelligence, bridging divide between high-dimensional sensory inputs action.

Language: Английский

Citations

1
Beyond learnability: understanding human visual development with DNNs DOI
Lei Yuan

Trends in Cognitive Sciences, Journal Year: 2024, Volume and Issue: 28(7), P. 595 - 596

Published: May 17, 2024

Language: Английский

Citations

1
Shape-Biased Learning by Thinking Inside the Box DOI Creative Commons
Niklas Müller, Cees G. M. Snoek, Iris I.A. Groen

et al.

bioRxiv (Cold Spring Harbor Laboratory), Journal Year: 2024, Volume and Issue: unknown

Published: June 1, 2024

Abstract Deep Neural Networks (DNNs) may surpass human-level performance on vision tasks such as object recognition and detection, but their model behavior still differs from human in important ways. One prominent example of this difference, the main focus our paper, is that DNNs trained ImageNet exhibit an texture bias, while humans are consistently biased towards shape. DNN shape-bias can be increased by data augmentation, next to being computationally more expensive, augmentation a biologically implausible method creating texture-invariance. We present empirical study texture-shape-bias showcasing high texture-bias correlates with background-object ratio. In addition, tight bounding boxes images sub-stantially shape than models full images. Using custom dataset high-resolution, annotated scene images, we show (I) systematically varies training boxes, (II) removal global result commonly applied cropping during increases (III) negatively correlated test accuracy positively cue-conflict created using following trend humans. Overall, improved supervision signal better reflects visual features truly belong to-be-classified deep neural networks. Our results also imply simultaneous alignment both classification strategy not achieved default suggesting need for new assessments behavioural between

Language: Английский

Citations

0
Parallel development of object recognition in newborn chicks and deep neural networks DOI Creative Commons
Lalit Pandey,

Donsuk Lee,

Samantha M. W. Wood

et al.

PLoS Computational Biology, Journal Year: 2024, Volume and Issue: 20(12), P. e1012600 - e1012600

Published: Dec. 2, 2024

How do newborns learn to see? We propose that visual systems are space-time fitters, meaning development can be understood as a blind fitting process (akin evolution) in which gradually adapt the spatiotemporal data distributions newborn’s environment. To test whether is viable theory for learning how see, we performed parallel controlled-rearing experiments on newborn chicks and deep neural networks (DNNs), including CNNs transformers. First, raised impoverished environments containing single object, then simulated those video game engine. Second, recorded first-person images from agents moving through virtual animal chambers used train DNNs. Third, compared viewpoint-invariant object recognition performance of When DNNs received same diet (training data) chicks, models developed common skills chicks. time teaching signal—space-time fitters—also showed patterns successes failures across viewpoints Thus, animals. argue fitters serve formal scientific systems, providing image-computable studying see raw experiences.

Language: Английский

Citations

0
Spatial Relation Categorization in Infants and Deep Neural Networks DOI Open Access
Guy Davidson, A. Emin Orhan, Brenden M. Lake

et al.

Published: April 24, 2023

Spatial relations, such as above, below, between, and containment, are important mediators in children’s understanding of the world (Piaget, 1954). The development these relational categories infancy has been extensively studied (Quinn, 2003) yet little is known about their computational underpinnings. Using developmental tests, we examine extent to which deep neural networks, pretrained on a standard vision benchmark or egocentric video captured from one baby’s perspective, form categorical representations for visual stimuli depicting relations. Notably, networks did not receive any explicit training We then analyze whether recover similar patterns ones identified development, reproducing relative difficulty categorizing different spatial relations stimulus abstractions. find that evaluate tend many observed with simpler “above versus below” “between outside”, but struggle match findings related “containment”. identify factors choice model architecture, pretraining data, experimental design contribute patterns, highlight predictions made by our modeling results. Our results open door infants’ earliest categorization abilities modern machine learning tools demonstrate utility productivity this approach.

Language: Английский

Citations

0