Are Neurons Actually Collapsed? On the Fine-Grained Structure in Neural Representations

Authors: Yongyi Yang, Jacob Steinhardt, Wei Hu

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental To study the above question, we design experiments to manually create a mismatch between the intrinsic structure of the input distribution and the explicit labels provided for training in standard classification datasets and measure how the last-layer representations behave in response to our interventions.
Researcher Affiliation Academia 1University of Michigan, Ann Arbor, Michigan, USA 2UC Berkeley, Berkeley, California, USA.
Pseudocode No The paper includes mathematical proofs and derivations, but does not present any pseudocode or algorithm blocks.
Open Source Code No The paper does not provide any links to open-source code for the methodology described.
Open Datasets Yes As an example, if we train on CIFAR-10 using only 5 coarse-grained labels (by combining two classes into one superclass) until convergence, we can reconstruct the original 10-class labels from the learned representations via unsupervised clustering.
Dataset Splits No The paper mentions using CIFAR-10 and its test set but does not explicitly describe the training/validation/test dataset splits or proportions in the main text.
Hardware Specification No The paper does not specify the hardware (e.g., GPU, CPU models) used for running the experiments.
Software Dependencies No The paper does not list specific software dependencies with version numbers (e.g., Python, PyTorch versions).
Experiment Setup Yes Specifically, we choose the initial learning rate in {10 1, 10 2, 10 3} (we apply a standard learning rate decay schedule) and the weight-decay rate in {5 10 3, 5 10 4, 5 10 5} and run all 9 possible combinations of them.