Deep Neural Networks and Visuo-Semantic Models Explain Complementary Components of Human Ventral-Stream Representational Dynamics

Journal of Neuroscience, Journal Year: 2023, Volume and Issue: 43(10), P. 1731 - 1741

Published: Feb. 9, 2023

Deep neural networks (DNNs) are promising models of the cortical computations supporting human object recognition. However, despite their ability to explain a significant portion variance in data, agreement between and brain representational dynamics is far from perfect. We address this issue by asking which features currently unaccounted for time series estimated multiple areas ventral stream via source-reconstructed magnetoencephalography data acquired participants (nine females, six males) during viewing. focus on visuo-semantic models, consisting human-generated labels categories, beyond explanatory power DNNs alone. report gradual reversal relative importance DNN versus as ventral-stream representations unfold over space time. Although lower-level visual better explained starting early (at 66 ms after stimulus onset), higher-level best accounted later 146 onset). Among features, parts basic categories drive advantage DNNs. These results show that component unexplained structured can be readily nameable aspects objects. conclude current fail fully capture dynamic cortex suggest path toward more accurate computations. SIGNIFICANCE STATEMENT When we view objects such faces cars our environment, dynamically at millisecond scale. reflect support fast robust have emerged framework modeling these but cannot yet account dynamics. Using observers viewing, objects, 'eye', 'wheel', 'face', above findings humans may part rely different recognition provide guidelines model improvement.

Language: Английский

---

The developmental trajectory of object recognition robustness: Children are like small adults but unlike big deep neural networks

Journal of Vision, Journal Year: 2023, Volume and Issue: 23(7), P. 4 - 4

Published: July 6, 2023

In laboratory object recognition tasks based on undistorted photographs, both adult humans and deep neural networks (DNNs) perform close to ceiling. Unlike adults’, whose performance is robust against a wide range of image distortions, DNNs trained standard ImageNet (1.3M images) poorly distorted images. However, the last 2 years have seen impressive gains in DNN distortion robustness, predominantly achieved through ever-increasing large-scale datasets—orders magnitude larger than ImageNet. Although this simple brute-force approach very effective achieving human-level robustness DNNs, it raises question whether human too, simply due extensive experience with (distorted) visual input during childhood beyond. Here we investigate by comparing core 146 children (aged 4–15 years) adults DNNs. We find, first, that already 4- 6-year-olds show remarkable distortions outperform Second, estimated number images had been exposed their lifetime. Compared various children’s high requires relatively little data. Third, when recognizing objects, children—like but unlike DNNs—rely heavily shape not texture cues. Together our results suggest emerges early developmental trajectory unlikely result mere accumulation input. Even though current match regarding they seem rely different more data-hungry strategies do so.

Language: Английский

---

The coming decade of digital brain research: A vision for neuroscience at the intersection of technology and computing

Imaging Neuroscience, Journal Year: 2024, Volume and Issue: 2, P. 1 - 35

Published: April 1, 2024

Abstract In recent years, brain research has indisputably entered a new epoch, driven by substantial methodological advances and digitally enabled data integration modelling at multiple scales—from molecules to the whole brain. Major are emerging intersection of neuroscience with technology computing. This science combines high-quality research, across scales, culture multidisciplinary large-scale collaboration, translation into applications. As pioneered in Europe’s Human Brain Project (HBP), systematic approach will be essential for meeting coming decade’s pressing medical technological challenges. The aims this paper to: develop concept decade digital discuss community large, identify points convergence, derive therefrom scientific common goals; provide framework current future development EBRAINS, infrastructure resulting from HBP’s work; inform engage stakeholders, funding organisations institutions regarding research; address transformational potential comprehensive models artificial intelligence, including machine learning deep learning; outline collaborative that integrates reflection, dialogues, societal engagement on ethical opportunities challenges as part research.

Language: Английский

---

Task-driven neural network models predict neural dynamics of proprioception

Cell, Journal Year: 2024, Volume and Issue: 187(7), P. 1745 - 1761.e19

Published: March 1, 2024

Proprioception tells the brain state of body based on distributed sensory neurons. Yet, principles that govern proprioceptive processing are poorly understood. Here, we employ a task-driven modeling approach to investigate neural code neurons in cuneate nucleus (CN) and somatosensory cortex area 2 (S1). We simulated muscle spindle signals through musculoskeletal generated large-scale movement repertoire train networks 16 hypotheses, each representing different computational goals. found emerging, task-optimized internal representations generalize from synthetic data predict dynamics CN S1 primates. Computational tasks aim limb position velocity were best at predicting activity both areas. Since task optimization develops better during active than passive movements, postulate is top-down modulated goal-directed movements.

Language: Английский

---

High-performing neural network models of visual cortex benefit from high latent dimensionality

PLoS Computational Biology, Journal Year: 2024, Volume and Issue: 20(1), P. e1011792 - e1011792

Published: Jan. 10, 2024

Geometric descriptions of deep neural networks (DNNs) have the potential to uncover core representational principles computational models in neuroscience. Here we examined geometry DNN visual cortex by quantifying latent dimensionality their natural image representations. A popular view holds that optimal DNNs compress representations onto low-dimensional subspaces achieve invariance and robustness, which suggests better should lower dimensional geometries. Surprisingly, found a strong trend opposite direction-neural with high-dimensional tended generalization performance when predicting cortical responses held-out stimuli both monkey electrophysiology human fMRI data. Moreover, high was associated learning new categories stimuli, suggesting higher are suited generalize beyond training domains. These findings suggest general principle whereby confers benefits cortex.

Language: Английский

---