Back

Brain-like variability in convolutional neural networks reveals evidence-, uncertainty- and bias-driven decision-making

Nakuci, J.

2026-01-12 neuroscience
10.1101/2025.09.16.676625 bioRxiv
Show abstract

Deep neural networks can appear confident even when weakly supported by inputs. Current approaches such as calibration and feature attribution relate confidence to error or map predictions to features, but do not quantify how the decision margin is supported. Here we introduce a layer-resolved Bias Dominance Index (BDI) that decomposes the margin into feature and bias components to localize where bias-dominance emerges. Across models - convolutional neural networks (CNN) and vision transformers (ViT)- BDI increases as feature support weakens and high confidence can coexist with bias-dominant decisions. Moreover, bias-dominance emerges in convolutional layers in CNNs and normalization layers in ViT. Perturbation analyses further show that bias components can be beneficial when readout weights are degraded. Finally, we operationalize decision composition as a two-parameter acceptance rule that combines confidence with BDI for auditing. Together, these results position BDI as a diagnostic for monitoring and triage, distinguishing feature-versus bias-driven decisions.

Matching journals

The top 3 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.