Self-supervision deep learning models are better models of human high-level visual cortex: The roles of multi-modality and dataset training size

bioRxiv (Cold Spring Harbor Laboratory), Год журнала: 2025, Номер unknown

Опубликована: Янв. 13, 2025

Abstract With the rapid development of Artificial Neural Network based visual models, many studies have shown that these models show unprecedented potence in predicting neural responses to images cortex. Lately, advances computer vision introduced self-supervised where a model is trained using supervision from natural properties training set. This has led examination their prediction performance, which revealed better than supervised for with language or image-only supervision. In this work, we delve deeper into models’ ability explain representations object categories. We compare differed objectives examine they diverge predict fMRI and MEG recordings while participants are presented different Results both self-supervision was advantageous comparison classification training. addition, predictor later stages perception, shows consistent advantage over longer duration, beginning 80ms after exposure. Examination effect data size large dataset did not necessarily improve predictions, particular models. Finally, correspondence hierarchy each cortex showed image only conclude consistently recordings, type reveals property activity, language-supervision explaining onsets, explains long very early latencies response, naturally sharing corresponding hierarchical structure as brain.

Язык: Английский

---

Telomeric SUMO level influences the choices of APB formation pathways and ALT efficiency

bioRxiv (Cold Spring Harbor Laboratory), Год журнала: 2025, Номер unknown

Опубликована: Янв. 20, 2025

Abstract Many cancers use an alternative lengthening of telomeres (ALT) pathway for telomere maintenance. ALT telomeric DNA synthesis occurs in telomere-associated PML bodies (APBs). However, the mechanisms by which APBs form are not well understood. Here, we monitored formation with time-lapse imaging employing CRISPR knock-in to track promyelocytic leukemia (PML) protein at endogenous levels. We found via two pathways: recruit proteins nucleate de novo, or fuse preformed bodies. Both nucleation and fusion require interactions between SUMO interaction motifs (SIMs). Moreover, APB is associated higher levels SUMOs SUMO-mediated recruitment helicase BLM, resulting more robust synthesis. Finally, further boosting enhances nucleation, BLM enrichment, Thus, high promote stronger activity.

Язык: Английский

---

Neural responses in early, but not late, visual cortex are well predicted by random-weight CNNs with sufficient model complexity

bioRxiv (Cold Spring Harbor Laboratory), Год журнала: 2025, Номер unknown

Опубликована: Фев. 6, 2025

ABSTRACT Convolutional neural networks (CNNs) were inspired by the organization of primate visual system, and in turn have become effective models cortex, allowing for accurate predictions stimulus responses. While training CNNs on brain-relevant object-recognition tasks may be an important pre-requisite to predict brain activity, CNN’s brain-like architecture alone already allow prediction activity. Here, we evaluated performance both task-optimized brain-optimized convolutional predicting responses across performed systematic architectural manipulations comparisons between trained untrained feature extractors reveal key structural components influencing model performance. For human monkey area V1, random-weight employing ReLU activation function, combined with either average or max pooling, significantly outperformed other functions. Random- weight matched their counterparts V1 The extent which can predicted correlated strongly network’s complexity, reflects non-linearity functions pooling operations. However, this correlation encoding complexity was weaker higher areas that are classically associated object recognition, such as IT. To test whether difference functional differences, network texture discrimination recognition tasks. Consistent our hypothesis, more than recognition. Our findings indicate sufficient comparable activity CNNs, while require precise configurations acquired through via gradient descent.

Язык: Английский

---

Language modulates vision: Evidence from neural networks and human brain-lesion models

Research Square (Research Square), Год журнала: 2025, Номер unknown

Опубликована: Март 27, 2025

Comparing information structures in between deep neural networks (DNNs) and the human brain has become a key method for exploring their similarities differences. Recent research shown better alignment of vision-language DNN models, such as CLIP, with activity ventral occipitotemporal cortex (VOTC) than earlier vision supporting idea that language modulates visual perception. However, interpreting results from comparisons is inherently limited due to "black box" nature DNNs. To address this, we combined model–brain fitness analyses lesion data examine how disrupting communication pathway systems causally affects ability vision–language DNNs explain VOTC. Across four diverse datasets, CLIP consistently outperformed both label-supervised (ResNet) unsupervised (MoCo) models predicting VOTC activity. This advantage was left-lateralized, aligning network. Analyses 33 stroke patients revealed reduced white matter integrity region left angular gyrus correlated decreased performance increased MoCo performance, indicating dynamic influence processing on These findings support integration modulation neurocognitive vision, reinforcing concepts models. The sensitivity similarity specific lesions demonstrates leveraging manipulation promising framework evaluating developing brain-like computer

Язык: Английский

---

Disentangling signal and noise in neural responses through generative modeling

bioRxiv (Cold Spring Harbor Laboratory), Год журнала: 2024, Номер unknown

Опубликована: Апрель 27, 2024

Abstract Measurements of neural responses to identically repeated experimental events often exhibit large amounts variability. This noise is distinct from signal , operationally defined as the average expected response across trials for each given event. Accurately distinguishing important, a target that worthy study (many believe reflects important aspects brain function) and it not confuse one other. Here, we describe principled modeling approach in which measurements are explicitly modeled sum samples multivariate distributions. In our proposed method—termed Generative Modeling Signal Noise (GSN)—the distribution estimated by subtracting data distribution. Importantly, GSN improves estimates distribution, but does provide improved individual events. We validate using ground-truth simulations show compares favorably with related methods. also demonstrate application empirical fMRI illustrate simple consequence GSN: disentangling components responses, denoises principal analysis dimensionality. end discussing other situations may benefit GSN’s characterization noise, such estimation ceilings computational models activity. A code toolbox provided both MATLAB Python implementations.

Язык: Английский

---

Universality of representation in biological and artificial neural networks

bioRxiv (Cold Spring Harbor Laboratory), Год журнала: 2024, Номер unknown

Опубликована: Дек. 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.

Язык: Английский

---

Fast and robust visual object recognition in young children

bioRxiv (Cold Spring Harbor Laboratory), Год журнала: 2024, Номер unknown

Опубликована: Окт. 16, 2024

Abstract By adulthood, humans rapidly identify objects from sparse visual displays and across significant disruptions to their appearance. What are the minimal conditions needed achieve robust recognition abilities when might these develop? To answer questions, we investigated upper-limits of children’s object abilities. We found that children as young 3 years successfully identified at speeds 100 ms (both forward backward masked) under disrupted viewing conditions. contrast, a range computational models implemented with biologically informed properties or optimized for did not reach child-level performance. Models only matched if they received more examples than capable experiencing. These findings highlight robustness human system in absence extensive experience important developmental constraints building plausible machines. Teaser The preschool rival those state-of-the-art artificial intelligence models.

Язык: Английский

---