One-Shot Pruning of Recurrent Neural Networks by Jacobian Spectrum Evaluation

Authors: Shunshi Zhang, Bradly C. Stadie

ICLR 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate on sequential MNIST, Billion Words, and Wikitext. Our method is evaluated with a GRU network on sequential MNIST, Wikitext, and Billion Words.
Researcher Affiliation Academia Matthew Shunshi Zhang University of Toronto matthew.zhang@mail.utoronto.ca Bradly C. Stadie Vector Institute
Pseudocode Yes Algorithm 1 Pruning Recurrent Networks
Open Source Code No The paper does not contain any statement about releasing source code for the described methodology or a link to a code repository.
Open Datasets Yes We evaluate on sequential MNIST, Billion Words, and Wikitext. (Lee et al., 2018) (Chelba et al., 2013)
Dataset Splits Yes We report the training and validation perplexities on a random 1% sample of the training set in Table 4. Table 1: Validation Error % of Various 400 Unit RNN Architectures after 50 Epochs of Training on Seq. MNIST
Hardware Specification Yes We trained all networks with a single Nvidia P100 GPU.
Software Dependencies No The paper mentions the use of the Adam optimizer and Glorot initialization, but does not specify version numbers for any programming languages or software libraries used in the experiments.
Experiment Setup Yes We use a minibatch size of 64 samples during training, and optimize using the Ada M optimizer (Kingma & Ba, 2014) with a learning rate of 1e-3. We use an initial hidden state of zeros for all experiments.