Untrained CNNs match human visual cortex at V1, research finds

A new study published on arXiv investigates how different learning rules in neural networks compare to human brain activity in visual processing. Researchers found that for early visual areas like V1 and V2, the network's architecture was more influential than the learning rule used. However, as the processing moved to higher-level areas like the Lateral Occipital Complex (LOC) and Inferior Temporal (IT) cortex, the learning rules began to show differentiation, with backpropagation proving more effective at LOC. AI

影响 Suggests early visual processing alignment in AI is primarily architecture-driven, with learning rules only differentiating at later stages.

排序理由 Academic paper comparing different AI learning rules against human fMRI data.

在 arXiv cs.LG 阅读 →

• arXiv
• backpropagation
• feedback alignment
• Nils Leutenegger
• predictive coding
• spike-timing-dependent plasticity
• THINGS-fMRI dataset
• V1
• V2

• 论文
• 其他

AI 生成摘要 · Google Gemini · 来自 1 个来源。 我们如何撰写摘要 →