An untrained CNN matches backpropagation at aligning with…
0 reactions
An untrained CNN matches backpropagation at aligning with human V1 — architecture matters more than learning for early visual cortex
New preprint comparing how different learning rules (backprop, feedback alignment, predictive coding, STDP) affect alignment with human visual cortex, measured with fMRI and RSA.
The most striking result: a CNN with completely random weights matches a fully trained backprop network at V1 and V2. The convolutional architecture alone produces representations that correlate with early visual cortex about as well as a trained model does.
Learning rules start to matter at higher visual areas (IT cortex), where backprop leads and predictive coding comes close using only biologically plausible local updates. Feedback alignment, often proposed as a bio-plausible alternative to backprop, actually makes representations worse than random.
Preprint: https://arxiv.org/abs/2604.16875
#technology
New preprint comparing how different learning rules (backprop, feedback alignment, predictive coding, STDP) affect alignment with human visual cortex, measured with fMRI and RSA.
The most striking result: a CNN with completely random weights matches a fully trained backprop network at V1 and V2. The convolutional architecture alone produces representations that correlate with early visual cortex about as well as a trained model does.
Learning rules start to matter at higher visual areas (IT cortex), where backprop leads and predictive coding comes close using only biologically plausible local updates. Feedback alignment, often proposed as a bio-plausible alternative to backprop, actually makes representations worse than random.
Preprint: https://arxiv.org/abs/2604.16875
#technology
| 2,000 mlx | |
| 3 | |
| 0 |