Keynotes | K&Ts | GACs | Talks | Posters | Search
Poster A in Poster Session A: Tuesday, August 4, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms
Circuit motifs underlying representational untangling
Jeffery W. Andrade1, Parisa A. Vaziri1, Talia Konkle1, George A. Alvarez1; 1Harvard University
Presenter: Jeffery W. Andrade
The ventral stream maps patterns of light to increasingly untangled, high-level category representations through hierarchical processing. What circuitry supports this untangling? Here we address this question in a fully observable “model organism,” AlexNet, by extracting the minimal inter-layer wiring sufficient to preserve accurate object categorization. Building on importance-based circuit pruning Hamblin et al., 2022, we refine logit-targeted pruning with (i) sign-aware pruning, which prunes excitatory and inhibitory connections on interleaved schedules to stabilize selection, and (ii) a cross-layer marginal-loss rule that allocates pruning across layers according to the increase in a calibrated one-vs-all loss per removed connection. Across 1,000 ImageNet categories, the resulting decision-preserving circuits retain only 4.5% of connections and are, on average, 5.7% below the calibrated unpruned baseline on corrected one-vs-all accuracy. Circuit structure reveals a shared macro-motif: dense early connectivity, progressively sparser convolutional stages, and dramatic sparsity in the fully connected readout. Cross-category wiring overlap also declines across the hierarchy, yet late-stage overlap remains significantly more structured than chance given matched circuit sparsity. At the superordinate level, the animate and inanimate domains contain substantially more shared computational routes within domain than the random a priori expectation, with increasingly separable circuitry along the hierarchy. Broadly, this method exposes the circuit-level computational structure underlying representational untangling, revealing systematic routing motifs across the hierarchy, and providing a computational approach for mapping category-computing circuitry across deep neural network systems.
Topic Area: Computational Models of Vision & Visual Cortex