Search Papers | Poster Sessions | All Posters
Poster A138 in Poster Session A - Tuesday, August 6, 2024, 4:15 – 6:15 pm, Johnson Ice Rink
Representational subspaces with different levels of abstraction in transformers
Brian Robinson1 (), Colin Conwell2, Michael Bonner2; 1Johns Hopkins University Applied Physics Laboratory, 2Johns Hopkins University
A widespread assumption in analyzing the representations of artificial neural networks (ANNs) and the brain is that neurons in the same ANN layer or cortical region have a shared level of abstraction. In this work, by analyzing the learned LayerNorm weights across a range of transformer networks, we find evidence for distinct subspaces in the network dimensions.In an in-depth analysis for a single vision transformer, we find three representational subspaces within each layer that can be identified by LayerNorm weights. In comparisons to human fMRI representations, we find distinct properties of these sub-spaces with two of the subspaces demonstrating higher representational similarity to early and late regions of the cortical visual hierarchy. These findings show that analyses of hierarchical feature processing in ANNs need to consider the role of subspaces with distinct representational properties.
Keywords: Vision Transformers Neural Network Representations Models of Human Vision fMRI