CAOS 2025 - 8 | Rovereto, May 7-9 | Talia Konkle

Value of the Information & Strength of the Argument

The talk provides valuable insights into the use of deep neural networks as model organisms for understanding visual processing. Konkle presents a clear argument for contrastive learning as a plausible objective for the visual system, supported by multiple empirical studies showing that such models achieve high brain predictivity. She also discusses important nuances, such as the role of visual diet and the emergence of category-selective units without explicit labels. The argumentation is solid, with references to specific experiments and data, though some claims are speculative and the framework is presented as one perspective among others.

Scientific Rigor, Source Quality, Title Accuracy

The talk demonstrates scientific rigor by grounding claims in published research and ongoing studies. Konkle cites specific papers (e.g., Wu et al. 2018, Wang & Isola, Zimmermann et al.) and presents data from her own lab. The title accurately reflects the content, focusing on mapping mechanisms to competencies. The talk is part of a scientific workshop, and the speaker is a recognized expert, contributing to its credibility. However, as a conference talk, it lacks the detail of a peer-reviewed paper, and some interpretations are open to debate.

The title accurately reflects the content: the speaker discusses mapping mechanisms to competencies using deep neural networks as model organisms for vision science.

8/10

The talk is given by a recognized expert in cognitive neuroscience, presenting a coherent framework supported by multiple empirical studies and references to published work. The claims are grounded in specific experiments and data, though some interpretations are speculative and the presentation is a conference talk rather than a peer-reviewed publication.

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

• 02:00Introduction: challenge of vision and the role of deep neural networks.
• 10:00Contrast between models as neural models vs. model organisms.
• 20:00Introduction to contrastive learning and instance-prototype contrastive learning.
• 30:00Results on brain predictivity of contrastive models vs. supervised models.
• 40:00Implications: high-fidelity perceptual interface and fine-grained discriminators.
• 50:00Role of visual diet and emergence of category-selective units.
• 1:00:00Circuit routing and sparse codes: pruning and lesioning studies.
• 1:10:00Active sensing and the role of goals in visual learning.
• 1:20:00Conclusion: visual system as a high-fidelity perceptual interface.

• CAOS 2025 Workshop — Official workshop page providing context for the talk.

• Wu et al. 2018 - Instance Discrimination — Paper that introduced instance-level contrastive learning, foundational to the talk.
• Wang & Isola - Understanding Contrastive Learning — Theoretical analysis of contrastive learning, cited in the talk.
• Zimmermann et al. - Contrastive Learning Inverts the Data Generating Process — Theoretical work on contrastive learning, mentioned in the talk.

• Bowers et al. - Deep Problems with Neural Network Models of Human Vision — This paper critiques the use of DNNs as models of human vision, arguing that they fail to capture key aspects of human perception, which contrasts with the optimistic view presented in the talk.