Projection Head is Secretly an Information Bottleneck

Authors: Zhuo Ouyang, Kaiwen Hu, Qi Zhang, Yifei Wang, Yisen Wang

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, our methods exhibit consistent improvement in the downstream performance across various real-world datasets, including CIFAR-10, CIFAR-100, and Image Net-100. We believe our theoretical understanding on the role of the projection head will inspire more principled and advanced designs in this field. Code is available at https://github.com/PKU-ML/Projector_Theory.
Researcher Affiliation Academia 1 College of Engineering, Peking University 2 School of EECS, Peking University 3 State Key Lab of General Artificial Intelligence, School of Intelligence Science and Technology, Peking University 4 MIT CSAIL 5 Institute for Artificial Intelligence, Peking University
Pseudocode No The paper describes methods and theoretical analyses using mathematical formulations, but it does not contain any clearly labeled pseudocode or algorithm blocks with structured steps.
Open Source Code Yes Code is available at https://github.com/PKU-ML/Projector_Theory.
Open Datasets Yes Empirically, our methods exhibit consistent improvement in the downstream performance across various real-world datasets, including CIFAR-10, CIFAR-100, and Image Net-100.
Dataset Splits Yes Empirically, our methods exhibit consistent improvement in the downstream performance across various real-world datasets, including CIFAR-10, CIFAR-100, and Image Net-100. ... We conduct our experiments on CIFAR-10, CIFAR-100, and Image Net-100, with Res Net-18 as our backbone.
Hardware Specification Yes All experiments are conducted with at most two NVIDIA RTX 3090 GPUs.
Software Dependencies No The paper does not explicitly mention specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes On CIFAR-10 and CIFAR-100, we use learning rate 0.4, weight decay 10 4, Info NCE temperature 0.2, and set λ to 10 4. Our projector adopts a Linear-Re LU-Linear structure, where we use 2048 as the hidden dimension and 256 as the output dimension. On Image Net-100, we use learning rate 0.3, weight decay 10 4, Info NCE temperature 0.2 , and set λ to 0.01. We use the same projector structure but change the hidden dimension to 4096 and the output dimension to 512.