Backpropagation destroys V1 brain alignment in one epoch, tracking RSA alignment to fMRI across training for BP, FA, predictive coding, and STDP [R]
Signal
78
Hype
15
En 3 lignesÉtude comparative de règles d'apprentissage (backprop, feedback alignment, predictive coding, STDP) via alignement RSA avec fMRI V1 humain. Backprop détruit 90% de l'alignement V1 après 1 epoch (r: 0.102→0.011), tandis que PC et STDP ne perdent que 25-31%. À epoch 40: PC/STDP >> BP/FA. Suggère un trade-off fondamental entre signaux d'erreur globaux (couches hautes) et alignement précoce.Lire la source
Ton avis ?
Résumé généré par Claude — vérifié par l'humain