Biologically grounded locality priors close the data gap for vision transformers in neural prediction

CiRF-ViT, a circular receptive-field vision transformer for mouse and salamander retinas, achieves the highest mean explained variance—0.94 (mouse) and 0.95 (salamander)—with only a few thousand stimulus-response pairs.

• Why it matters: Transformers are expected to underperform when neural prediction has low data, where CNNs and Gaussian processes are often favored; this work tests whether an inductive bias can close that gap.
• What they did: The model applies a differentiable per-neuron circular crop in feature space, centered on each neuron's receptive field with a per-neuron radius, so attention cost scales with receptive-field size; it is evaluated on 41 mouse ganglion cells and 49 salamander ganglion cells.
• The result: Compared with CNN and Gaussian-process baselines, CiRF-ViT reaches the top mean explained variance on both datasets and, via zero-shot LSTA probing, reproduces polarity inversion that a linear model cannot capture.

AI summary of the abstract · may contain errors · not medical advice · read the paper.