Coreset-driven sample selection for fMRI distillation data

[OnePunchMonk]

Idea

tiny-tribe distills a 4.7B-parameter TRIBE v2 teacher into a ~14M-param student for predicting fMRI cortical responses to naturalistic video. Currently (as far as the repo describes) this trains on the full available video/fMRI dataset. Apply coreset selection (per PixCoreset / Coresets-for-Image-Segmentation) to pick the most informative subset of training samples for the distillation objective instead.

Hypothesis

A coreset-selected subset of naturalistic video/fMRI pairs — chosen to maximize coverage of the teacher's response-diversity — should get most of the way to full-dataset distillation accuracy at a fraction of the data/compute, the same way coreset sampling preserved segmentation quality with much less data in your image work.

Proposed scope (v0)

1. Define a coreset selection objective over the distillation training set — likely based on diversity/coverage in teacher-embedding space (e.g. k-center or gradient-based coreset selection) rather than raw pixel/signal similarity.
2. Train the 14M student on: (a) full dataset, (b) random subsample at matched size, (c) coreset-selected subsample at matched size.
3. Evaluate all three on held-out cortical-response prediction accuracy (whatever metric tiny-tribe already reports — likely voxel-wise correlation or similar).
4. Report the data-reduction curve: how much of the dataset can be dropped via coreset selection before accuracy on held-out responses meaningfully degrades, vs. the same curve for random subsampling.

Why this is worth doing

• Directly reuses coreset methodology you've already validated in vision (two prior repos) on a new, higher-stakes domain (neuroscience data is expensive/scarce to collect, unlike web-scale image data) — the value proposition of coresets is strongest exactly where data collection is the bottleneck.
• If the result holds, it's a genuinely useful finding for anyone doing brain-encoding model distillation, not just a repeat of an existing technique in a new domain.

Portfolio framing

This connects two of your existing research threads (coresets, neuro-AI distillation) rather than starting a new one — the report itself (the reduction curve, random vs. coreset) is the artifact, more than the code.