Scale-Invariance in AI Representation Predicts AI-Brain Alignment

Junjie Yu, Wenxiao Ma, Chen Wei, Jianyu Zhang, Haotian Deng, Zihan Deng, Yi Guo, Quanying Liu

Keywords: AI-Brain Alignment, fMRI, intrinsic dimension, pre-trained model

Abstract: Understanding why some neural network representations align better with brain activity is essential for uncovering neural coding principles and developing human-like AI. While prior work has largely focused on model-level factors, such as dataset scale and task design, we focus on the rarely explored, yet more in-depth embedding level. Motivated by evidence that scale-invariance is widespread in biological neural systems, we identify it as a key embedding-level property. Analyzing 60 pretrained visual models and fMRI responses to natural images, we find that embeddings with stronger scale-invariance align better with fMRI. Training strategies modulate scale-invariance, with larger pretraining datasets enhancing it and fine-tuning reducing it, thereby affecting alignment performance. These findings establish scale-invariance as a fundamental embedding-level property that links training strategies to brain-like representations and suggest its potential as a guiding principle for designing more human-like AI.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 8782