iiinova — where your digital self keeps evolving

Erin Grant @eringrant.me · 07/12/2025

@dataonbrainmind.bsky.social starting now in Room 10 with opening remarks from @crji.bsky.social and the first invited talk from @dyamins.bsky.social!

0113

Erin Grant @eringrant.me · 13/08/2025

Function-representation dissociations and the representation-computation link persist in deep nonlinear networks! Using function-invariant reparametrisations (@bsimsek.bsky.social), we break representational identifiability but degrade generalization (a computational consequence).

Function-representation dissociation in ReLU networks. (A-B) MNIST representations before/after prediction-preserving reparametrisation. (C) RSM after function-preserving reparametrisation. (D-E) Performance under input/parameter noise for different solution types.

110

Erin Grant @eringrant.me · 13/08/2025

We demonstrate that representation analysis and comparison is ill-posed, giving both false negatives and false positives, unless we work with *task-specific representations*. These are interpretable *and* robust to noise (i.e., representational identifiability comes with computational advantages).

Hidden-layer representations for a semantic hierarchy task. (A) Task structure. (B) Input/target encoding. (C-E) Hidden representations and representational similarity matrices for task-agnostic (C: LSS) vs. task-specific (D: MRNS, E: MWNS) solutions.

131

Erin Grant @eringrant.me · 13/08/2025

We parametrised this solution hierarchy to find differences in handling of task-irrelevant dimensions: Some solutions compress away (creating task-specific, interpretable representations), while others preserve arbitrary structure in null spaces (creating arbitrary, uninterpretable representations).

The solution manifold. (A) Solution manifold for a 3-parameter linear network, showing GLS and constrained LSS, MRNS, and MWNS solutions. (B-E) Input/output weight relationships and parametrisation structure for each solution type.

120

Erin Grant @eringrant.me · 13/08/2025

To analyse this dissociation in a tractable model of representation learning, we characterize *all* task solutions for two-layer linear networks. Within this solution manifold, we identify a solution hierarchy in terms of what implicit objectives are minimized (in addition to the task objective).

Task solution hierarchy defined by implicit regularisation objectives.

140

Erin Grant @eringrant.me · 13/08/2025

Deep networks have parameter symmetries, so we can walk through solution space, changing all weights and representations, while keeping output fixed. In the worst case, function and representation are *dissociated*. (Networks can have the same function with the same or different representation.)

Example of a failure case. (A) A random walk on the solution manifold of a two-layer linear network reveals that weights can change continuously, inducing changes in the (B) network parametrisation and thus the (C) hidden-layer representations, while preserving the (D) network output.

130

Erin Grant @eringrant.me · 13/08/2025

This GAC focuses on three debates/questions around benchmarks in cognitive science (the what, why, and how): (1) Should data or theory come first? (2) Should we focus on replication or exploration? (3) What incentives should we build up, if we choose to invest effort as a community?

The three questions of the GAC?
1. What should benchmarks measure?
2. What should the goals of a benchmark be?
3. How should benchmarks be structured?

140

Erin Grant @eringrant.me · 13/08/2025

Our #CCN2025 GAC debate w/ @gretatuckute.bsky.social, Gemma Roig (www.cvai.cs.uni-frankfurt.de), Jacqueline Gottlieb (gottlieblab.com), Klaus Oberauer, @mschrimpf.bsky.social &‬ @brittawestner.bsky.social asks: 📊 What benchmarks are useful for cognitive science? 💭 2025.ccneuro.org/gac

Speakers and organizers of the GAC debate. Time and location of the GAC debate: 5 PM in Room C1.03.

15016

Erin Grant @eringrant.me · 13/12/2024

If you missed it at the #NeurIPS2024 posters! Work led by @leonlufkin.bsky.social on analytical dynamics of localization in simple neural nets, as seen in real+artificial nets and distilled by @aingrosso.bsky.social @sebgoldt.bsky.social. Leon is a fantastic collaborator + looking for PhD positions!

Leon presenting poster at the NeurIPS poster session to an intrigued audience

1223