#2 · ML / AI methods

Interpretability Tools Recover Known Visual Selectivity in Neural Models

The dominant thread is validation of machine learning models of neural activity against established neurophysiology. Two items build or benchmark models of neural data directly: a spatiotemporal TDANN with contrastive training reproduces macaque MT direction maps and pinwheel structure, and SpikeProphecy provides a causal autoregressive benchmark for spike-count forecasting across 105 Neuropixels sessions. Other items probe trained encoders or foundation models with attribution and visualization methods. Feature visualization recovers the V1-V4 hierarchy and FFA, MT, and PPA selectivity in TRIBE v2; mechanistic analysis localizes image features driving human visual cortex voxels with causal validation; and LRP on EEG transformers exposes ocular shortcuts in motor imagery alongside a central arousal signature. What is new is the use of these checks as discovery tools, including a closed-loop framework that converts digital-twin models of macaque V1 and V4 neurons into verifiable language descriptions of selectivity.