The geometry of the hidden representation of large transformer models

Diego Doimo · Figshare · 2023

This dataset contains some statistics related to the hidden representations of the ImageNet training set on the Image-GPT transformer. More precisely, for each data point, it contains the nearest neighbors' distance ratios to compute the intrinsic dimension of the hidden representation and the indices of the nearest neighbor data points to compute the overlap with the classification labels.

Read the paper · More papers on PaperTik