Artificial Neural Network Language Models Predict Human Brain Responses to Language Even After a Developmentally Realistic Amount of Training

Neurobiology of Language, Journal Year: 2024, Volume and Issue: 5(1), P. 43 - 63

Published: Jan. 1, 2024

Abstract Artificial neural networks have emerged as computationally plausible models of human language processing. A major criticism these is that the amount training data they receive far exceeds humans during learning. Here, we use two complementary approaches to ask how models’ ability capture fMRI responses sentences affected by data. First, evaluate GPT-2 trained on 1 million, 10 100 or billion words against an benchmark. We consider 100-million-word model be developmentally in terms given this similar what children are estimated exposed first years life. Second, test performance a 9-billion-token dataset reach state-of-the-art next-word prediction benchmark at different stages training. Across both approaches, find (i) already achieve near-maximal capturing sentences. Further, (ii) lower perplexity—a measure performance—is associated with stronger alignment data, suggesting received enough sufficiently high also acquire representations predictive responses. In tandem, findings establish although some necessary for ability, realistic (∼100 million words) may suffice.

Language: Английский

---

Language in Brains, Minds, and Machines

Annual Review of Neuroscience, Journal Year: 2024, Volume and Issue: 47(1), P. 277 - 301

Published: April 26, 2024

It has long been argued that only humans could produce and understand language. But now, for the first time, artificial language models (LMs) achieve this feat. Here we survey new purchase LMs are providing on question of how is implemented in brain. We discuss why, a priori, might be expected to share similarities with human system. then summarize evidence represent linguistic information similarly enough enable relatively accurate brain encoding decoding during processing. Finally, examine which LM properties—their architecture, task performance, or training—are critical capturing neural responses review studies using as silico model organisms testing hypotheses about These ongoing investigations bring us closer understanding representations processes underlie our ability comprehend sentences express thoughts

Language: Английский

---

Animal models of the human brain: Successes, limitations, and alternatives

Current Opinion in Neurobiology, Journal Year: 2025, Volume and Issue: 90, P. 102969 - 102969

Published: Feb. 1, 2025

Language: Английский

---

Semantic Processing of Argument Structure during Naturalistic Story Listening: Evidence from Computational Modeling on fMRI

NeuroImage, Journal Year: 2025, Volume and Issue: unknown, P. 121253 - 121253

Published: May 1, 2025

Language: Английский

---

Machine learning in seismic structural design: an exploration of ANN and tabu-search optimization

Asian Journal of Civil Engineering, Journal Year: 2023, Volume and Issue: 25(3), P. 2367 - 2377

Published: Nov. 4, 2023

Language: Английский

---

Modeling dynamic social vision highlights gaps between deep learning and humans

Published: June 11, 2024

Deep learning models trained on computer vision tasks are widely considered the most successful of human to date. The majority work that supports this idea evaluates how accurately these predict brain and behavioral responses static images objects natural scenes. Real-world vision, however, is highly dynamic, far less has focused evaluating accuracy deep in predicting stimuli move, involve more complicated, higher-order phenomena like social interactions. Here, we present a dataset videos captions involving complex multi-agent interactions, benchmark 350+ image, video, language neural videos. As with prior work, find many reach noise ceiling visual scene features along ventral stream (often primary substrate object recognition). In contrast, image poorly action interaction ratings lateral (a pathway increasingly theorized as specializing vision). Language (given sentence videos) better than either or video models, but they still perform at stream. Together results identify major gap AI's ability match highlight importance studying contexts.

Language: Английский

---

Divergences in color perception between deep neural networks and humans

Cognition, Journal Year: 2023, Volume and Issue: 241, P. 105621 - 105621

Published: Sept. 14, 2023

Deep neural networks (DNNs) are increasingly proposed as models of human vision, bolstered by their impressive performance on image classification and object recognition tasks. Yet, the extent to which DNNs capture fundamental aspects vision such color perception remains unclear. Here, we develop novel experiments for evaluating perceptual coherence embeddings in DNNs, assess how well these algorithms predict similarity judgments collected via an online survey. We find that state-of-the-art DNN architectures - including convolutional transformers provide strikingly diverge from (i) images with controlled properties, (ii) generated searches, (iii) real-world canonical CIFAR-10 dataset. compare against interpretable cognitively plausible model based wavelet decomposition, inspired foundational theories computational neuroscience. While one deep learning a trained style transfer task captures some perception, our algorithm provides more coherent better compared all examine. These results hold when altering high-level visual used train similar (e.g., versus segmentation), examining different layers given architecture. findings break new ground effort analyze representations machine improve ability serve vision. Implications learning, embodied cognition discussed.

Language: Английский

---

Information-Restricted Neural Language Models Reveal Different Brain Regions’ Sensitivity to Semantics, Syntax, and Context

Neurobiology of Language, Journal Year: 2023, Volume and Issue: 4(4), P. 611 - 636

Published: Jan. 1, 2023

A fundamental question in neurolinguistics concerns the brain regions involved syntactic and semantic processing during speech comprehension, both at lexical (word processing) supra-lexical levels (sentence discourse processing). To what extent are these separated or intertwined? address this question, we introduce a novel approach exploiting neural language models to generate high-dimensional feature sets that separately encode information. More precisely, train model, GloVe, GPT-2, on text corpus from which selectively removed either We then assess features derived information-restricted still able predict fMRI time courses of humans listening naturalistic text. Furthermore, determine windows integration processing, manipulate size contextual information provided GPT-2. The analyses show that, while most comprehension sensitive features, relative magnitudes effects vary across regions. Moreover, best fitted by more spatially dissociated left hemisphere than right one, shows sensitivity longer contexts left. novelty our lies ability control for encoded models' embeddings manipulating training set. These "information-restricted" complement previous studies used probe bases language, shed new light its spatial organization.

Language: Английский

---

Universality of representation in biological and artificial neural networks

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

Published: Dec. 26, 2024

Abstract Many artificial neural networks (ANNs) trained with ecologically plausible objectives on naturalistic data align behavior and representations in biological systems. Here, we show that this alignment is a consequence of convergence onto the same by high-performing ANNs brains. We developed method to identify stimuli systematically vary degree inter-model representation agreement. Across language vision, then showed from high-and low-agreement sets predictably modulated model-to-brain alignment. also examined which stimulus features distinguish high-from sentences images. Our results establish universality as core component provide new approach for using uncover structure computations.

Language: Английский

---

Can ChatGPT help researchers understand how the human brain handles language?

Proceedings of the National Academy of Sciences, Journal Year: 2024, Volume and Issue: 121(25)

Published: June 14, 2024

This study unravels a concerning capability in Large Language Models (LLMs): the ability to understand and induce deception strategies. As LLMs like GPT-4 intertwine with human communication, aligning them values becomes paramount. ...Large language models (LLMs) are currently at forefront of intertwining AI systems communication everyday life. Thus, is great importance. However, given steady increase reasoning abilities, ...

Language: Английский

---