CCN 2026 | K&T Keynotes

Value of the Information & Strength of the Argument

The keynote provides valuable insights into the limitations of representational comparisons, a core method in computational neuroscience. It systematically deconstructs assumptions underlying the task optimization framework, using simple linear and nonlinear networks to illustrate degeneracy. The argumentation is clear and logical, building from the problem of degeneracy to potential solutions. The speakers effectively highlight the tension between the neatness of models and the messiness of neural computation, and they propose a shift towards representational pluralism. The value lies in its critical perspective, which is essential for advancing the field, though it lacks concrete empirical examples and detailed methodological guidance.

Scientific Rigor, Source Quality, Title Accuracy

The keynote demonstrates scientific rigor by referencing established concepts such as the contravarian principle and the platonic representation hypothesis, and by acknowledging ongoing debates in the field. However, the talk does not provide explicit citations or references to specific studies, relying instead on general knowledge. The title accurately reflects the content, which is a keynote session introducing tutorials on representational comparisons. The session is well-structured, with clear transitions between speakers, and the content is consistent with the theme of the conference. The lack of detailed sources is a minor weakness, but the overall scientific quality is high.

The title accurately reflects the content, which is a keynote session introducing three tutorials on representational comparisons and related topics at CCN 2026.

7/10

The keynote presents a critical perspective on representational comparisons in neuroscience, drawing on established concepts and recent debates, but lacks detailed citations and empirical data in the talk itself.

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

• 00:00Introduction by moderator Angela Radulescu, introducing the keynote session and speakers.
• 02:00Erin Grant begins her talk on rethinking representational comparisons, introducing the concept of neural computation being messy and models being neat.
• 05:00Grant explains the paradigm of representational comparison using RSA, detailing the pipeline of collecting stimulus responses and computing similarity matrices.
• 10:00Grant discusses assumptions underlying task optimization, including the idea that function determines representation and that harder tasks eliminate degeneracies.
• 15:00Grant introduces the concept of degeneracy in biological and artificial neural networks, using rotational variability as an example.
• 20:00Grant discusses the tension between universality and idiosyncrasy, referencing debates in biology, cognitive science, and AI.
• 25:00Lucas Braun takes over, explaining the use of simple linear networks to show that function does not determine representation, using a semantic hierarchy example.
• 30:00Braun demonstrates that multiple representational geometries can implement the same function, and poses the question of why task-specific representations are often observed.
• 35:00Braun introduces the nonlinear case, mentioning rescaling symmetries and other degeneracies in nonlinear networks, and hints at conditions for unique geometries.
• 40:00Braun concludes by introducing the third part on representational pluralism, led by Eleanor Holton, and emphasizes the need to study diverse representations.

• CCN 2026 Keynotes & Tutorials — Official conference page for the keynote and tutorial sessions, providing details on the topics and organizers.

• CCN 2026 Keynotes & Tutorials — Official conference page, consistent with the topics discussed in the keynote.