Search Papers | Poster Sessions | All Posters
Poster A141 in Poster Session A - Tuesday, August 6, 2024, 4:15 – 6:15 pm, Johnson Ice Rink
High-dimensional alignment of neural networks and visual cortex
Tailai Shen1, Colin Conwell1, Michael F. Bonner1 (); 1Johns Hopkins University
Research into the representational similarity between deep neural networks (DNNs) and the human visual cortex aims to deepen our understanding of both systems. Here we explored the alignment between DNNs and the ventral visual stream by extending conventional representational similarity statistics to a spectrum of similarities across thousands of latent dimensions. The spectrum is generated by computing the correlations between aligned latent dimensions in model and brain representations. Using this approach, we found that DNN layers and regions of visual cortex have shared high-dimensional representations, spanning thousands of dimensions. The dimensionality of these shared representations exhibits an overall decrease from early to late visual regions. However, by separately reducing the channel and spatial dimensions of DNNs, we found that there is a complex relationship between dimensionality and the visual hierarchy. Specifically, in early visual regions, the alignment with DNNs relies heavily on high spatial dimensionality, whereas in late visual regions, it relies heavily on high channel dimensionality. Together, these results demonstrate the potential insights that can be gained by characterizing the full spectrum of high-dimensional alignment between computational models and visual cortex.
Keywords: vision deep neural network latent dimensions representational alignment