Contributed Talks: “Visual Processing in Brains and Models II” - CCN 2025

Item: Contributed Talks: “Visual Processing in Brains and Models II” - CCN 2025
Rating: 4
Author: PostSingularity Institute

by Cognitive Computational Neuroscience

Life & Natural Systems Neuroscience PSANNeurosciences PSAN5Cognitive and behavioural neuroscience

🎙 Cognitive Computational Neuroscience 👥 4K 📅 October 8, 2025 ⏱ 51 min 👁 233 📄 conference talks 🧭 2026-08-15

Available in: English (current) Français

Keywords

curvature perceptionpower lawout-of-distributionfMRIencoding models

Summary

This video is a recording of the contributed talks session ‘Visual Processing in Brains and Models II’ at the Cognitive Computational Neuroscience Conference 2025 in Amsterdam. It features five presentations. Laura Mai Stoinski discusses quantifying perceived curvature in natural object images, showing that perceived curvature is a reliable measure that correlates with brain organization in occipitotemporal cortex and can be predicted by a model based on image embeddings. Raj Magesh Gauthaman (presented by Mick Bonner) presents work on shared high-dimensional latent structure in neural and mental representations of objects, finding that both brain and behavioral representations exhibit power-law distributions of variance across dimensions, and that these dimensions correspond between brain and behavior. Alessandro Thomas Gifford introduces NSD-synthetic, a new fMRI dataset of synthetic images for out-of-distribution testing of vision models, demonstrating that current encoding models generalize poorly to out-of-distribution stimuli. Carmen Amme investigates encoding of fixation-specific visual information, finding no evidence of information carry-over between fixations. Wieger H. Scheurer presents a hierarchy of spatial predictions across human visual cortex during natural vision. The talks are followed by brief Q&A sessions. The session highlights recent advances in understanding visual processing through large-scale datasets and computational models.

195 words

Critical Evaluation

Value of the Information & Strength of the Argument

The talks provide valuable insights into visual neuroscience, combining large-scale neuroimaging data with computational modeling. Each presentation is well-structured, with clear hypotheses, methods, and results. The argumentation is solid, relying on empirical evidence and statistical analyses. The use of large datasets like THINGS and NSD strengthens the conclusions. The Q&A sessions add depth, addressing potential limitations and alternative interpretations.

Scientific Rigor, Source Quality, Title Accuracy

The scientific rigor is high, as the talks are based on peer-reviewed research presented at a major conference. The speakers reference relevant literature and datasets, and the methods are described in sufficient detail. The title accurately reflects the content, which is a session of contributed talks on visual processing. The video description lists the speakers and talk titles, providing clear attribution. No external sources are cited in the video itself, but the talks likely reference published papers, which are not explicitly listed in the description.

159 words

Title / Content Match

The title accurately describes the content: a session of contributed talks on visual processing in brains and models at CCN 2025.

Quality & Reliability

8/10

The video consists of five peer-reviewed contributed talks at a reputable conference (CCN 2025). Each talk presents original research with methodological details and references to published work. The content is scientifically rigorous, though limited by the brevity of conference presentations.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

THINGS dataset — Mentioned by Laura Stoinski and Mick Bonner as a large-scale stimulus database.
Natural Scenes Dataset (NSD) — Mentioned by Alessandro Gifford as a large-scale fMRI dataset.
MLV toolbox — Mentioned by Laura Stoinski for computing mid-level visual features.
ResNet-50 — Mentioned by Laura Stoinski and Alessandro Gifford as a deep neural network for image embeddings.
CLIP — Mentioned by Laura Stoinski as an alternative model for predicting perceived curvature.

Concurring Sources

THINGS dataset — Used by multiple speakers for behavioral and neural data.
Natural Scenes Dataset — Used by Alessandro Gifford for fMRI data.

Contribution & Novelties

The session presents several novel contributions: Laura Stoinski provides a new image-computable model of perceived curvature and demonstrates its relevance to brain organization. Raj Magesh Gauthaman reveals that mental representations of objects follow a power-law structure similar to neural representations, suggesting a fundamental principle. Alessandro Gifford introduces NSD-synthetic, a new dataset for out-of-distribution testing, which is crucial for evaluating model generalization. These contributions advance our understanding of visual processing and provide new tools for the community.

Pour aller plus loin :

THINGS dataset — Large-scale multimodal dataset for object recognition.
Natural Scenes Dataset — Large-scale fMRI dataset of natural images.
Power law in neural representations — General concept of power laws in natural phenomena.
Encoding models in neuroscience — Overview of encoding models used to predict brain responses.

128 words

Radar Profile

The radar profile shows high scores in quality and quantity of information, with moderate technical level and reliability. This reflects the advanced scientific content and the expertise of the speakers, though the format limits depth.

Reliability 8/10

💬 No comments were provided for analysis.