Do brain-like learning rules make brain-like vision models?
Training vision networks with backprop and with biologically plausible learning rules, then comparing each one to human EEG responses to the same images.
2026–27In progressEEGML
The question
Do local, biologically plausible learning rules explain the visual cortex better than standard backpropagation?
The plan
- Train the same convolutional neural network on the same image classification task three ways: standard backpropagation, feedback alignment and predictive coding.
- Extract each trained model’s layer activations.
- Compare those activation patterns with real human EEG responses to the same images.
Example
- Classify natural images from the THINGS image set.
- Build a representational dissimilarity matrix (RDM) for each model layer.
- Correlate each one against EEG-derived RDMs from the THINGS-EEG2 dataset, at each timepoint.
- Find which learning rule best predicts the time course of the brain’s visual response.
What it produces
- Classification accuracy for each learning rule
- A brain-similarity (RSA) score for each rule at each timepoint
- A comparison of compute and training cost
- A test of whether “biologically plausible” rules are actually more brain-like, not just less accurate
Dataset: THINGS-EEG2.