Supervised Training Rapidly Degrades Early Visual Cortex Alignment Across Biologically Plausible Learning Rules
arXiv:2605.30556v1 Announce Type: new Abstract: Random, untrained neural networks consistently match or exceed trained networks in representational similarity to early visual cortex. This puzzling finding challenges the assumption that learning improves brain alignment. We investigate it by tracking representational similarity analysis (RSA) alignment to human fMRI data across training for four learning rules: backpropagation (BP), feedback alignment (FA), predictive coding (PC), and spike-timing-dependent plasticity (STDP). Using 720 object images from the THINGS database and fMRI data from t
This research is published as AI capabilities rapidly advance, and understanding the biological alignment of AI models becomes crucial for developing more effective and brain-like intelligence.
The finding challenges the assumption that supervised learning invariably improves AI's biological plausibility, which could influence future AI research directions and the design of neuro-inspired architectures.
Our understanding of the 'why' behind AI model performance and its relationship to biological intelligence could shift, potentially leading to alternative training paradigms that prioritize alignment over raw performance metrics.
- · Neuroscience researchers
- · AI ethicists
- · Developers of foundational AI models
- · AI models without biological alignment as a design goal
- · Purely performance-driven supervised learning paradigms
This research suggests a re-evaluation of supervised learning's role in creating biologically aligned AI.
It could spur increased investment in unsupervised or self-supervised learning methods that might naturally lead to better brain alignment.
A deeper understanding of AI-brain alignment may lead to more energy-efficient and generalizable AI, potentially impacting the computational demands of future AI systems.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.LG