Backpropagation destroys V1 brain alignment 90% in one epoch, study finds
Backprop destroys V1 alignment in a single training epoch — PC and STDP barely budge.
A new paper in a series tracking learning rules vs. human fMRI (THINGS dataset, V1–IT, N=3 subjects) reveals a stark trade-off: backpropagation (BP) destroys 90% of V1 brain alignment in just one training epoch. Using RSA alignment measured at 8 checkpoints across 5 seeds, BP dropped from r=0.102 to r=0.011 (p=0.031), consistent across all seeds. Feedback alignment (FA) dropped 49%, while predictive coding (PC) and STDP dropped only 25-31% and stabilized. By epoch 40, PC (r=0.064) and STDP (r=0.059) significantly outperformed BP (r=0.022) and FA (r=0.019), with Cohen's d > 5.
The study suggests a fundamental trade-off: global error signals (BP, FA) build better higher-level representations (LOC showed +0.011 increase for BP) but destroy early visual cortex alignment, while local learning rules preserve it. The degradation rate tracks error signal globality: exact gradients > random feedback > local prediction errors. Limitations include 5 seeds capping resolution at p≈0.031, training on 32×32 CIFAR-10 evaluated on 224×224 THINGS (domain shift), and the LOC increase not tested for significance. The authors predict deeper models show the same pattern but more slowly.
- Backpropagation drops V1 alignment 90% after one training epoch (r=0.102→0.011, p=0.031, consistent across 5 seeds).
- Predictive coding and STDP lose only 25-31% of V1 alignment and stabilize, outperforming BP by Cohen's d > 5 at epoch 40.
- Global error signals (BP, FA) improve earlier cortex alignment (LOC) but destroy early V1 alignment; local learning rules preserve it.
Why It Matters
This reveals a fundamental trade-off in AI learning rules: global error signals help higher cortex but damage early visual representations.