Rare Gems: Finding Lottery Tickets at Initialization

Authors: Kartik Sreenivasan, Jy-yong Sohn, Liu Yang, Matthew Grinde, Alliot Nagle, Hongyi Wang, Eric Xing, Kangwook Lee, Dimitris Papailiopoulos

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we present the experimental results for the performance of GEM-MINER across various tasks.
Researcher Affiliation Collaboration c Carnegie Mellon University m Mohamed Bin Zayed University of Artificial Intelligence p Petuum, Inc. w University of Wisconsin-Madison
Pseudocode Yes Algorithm 1: GEM-MINER
Open Source Code Yes Our codebase can be found at https://github.com/ksreenivasan/pruning_is_enough.
Open Datasets Yes We evaluate our algorithm on (Task 1) CIFAR-10 classification... (Task 2) Tiny Image Net classification... (Task 3) Finetuning on the Caltech-101 [7] dataset... and (Task 4) CIFAR-100 classification...
Dataset Splits No The paper mentions training and testing but does not explicitly provide details about a validation dataset split (e.g., percentages or sample counts for a validation set) in the main text.
Hardware Specification No The paper does not provide specific details about the hardware used for running the experiments (e.g., CPU/GPU models, memory specifications).
Software Dependencies No The paper mentions some algorithms and optimizers used (e.g., Adam), but does not specify software dependencies with version numbers (e.g., Python 3.x, PyTorch 1.x).
Experiment Setup Yes If a network reaches its best accuracy after E epochs of dense training, then we run GEM-MINER for E epochs from random init to get a sparse subnetwork at initialization, and then run weight training on the sparse subnetwork for another E epochs. For CIFAR-10, Mobile Net-V2 experiments, where we apply GEM-MINER for 300 epochs and then finetune the sparse model for another 300 epochs, to reach 98.6% sparse model.