AI BriefingDiscovery of the Brain-Alignment Collapse Phenomenon in V1 Caused by Backpropagation
Key point
Backpropagation sharply reduces brain alignment in the V1 region immediately after training, while local learning rules such as predictive coding preserve it.
Details
Analyzing how well learning rules align with human fMRI data (THINGS dataset) via RSA (Representational Similarity Analysis), it was found that the globality of the error signal plays a decisive role in maintaining brain alignment.
Key Findings:
- Sharp Alignment Collapse from Backpropagation (BP): BP reduces brain alignment in the V1 region by over 90% after just 1 epoch (r: 0.102 → 0.011).
- Stability of Local Learning Rules: Predictive Coding (PC) and STDP show only a 25–31% decline in alignment, remaining stable.
- Trade-off Between Expressiveness and Alignment: While BP destroys V1 alignment, it slightly increases alignment in the LOC (Object-selective cortex) region. This suggests that a global error signal is advantageous for building high-level object representations, but harms alignment in early visual cortex.
In conclusion, a correlation was confirmed in which the more global the error signal (BP > FA > PC/STDP), the faster the degradation of brain alignment.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Sakana AI Releases 'Augmented Lagrangian Predictive Coding' Capable of Training 1000-Layer Neural NetworksTBC Partners with AWS to Commercialize Neuron-Based AI Video Model