Identifying Learning Rules From Neural Network Observables

Authors: Aran Nayebi, Sanjana Srivastava, Surya Ganguli, Daniel L. Yamins

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

Reproducibility Variable Result LLM Response
Research Type Experimental Simulating idealized neuroscience experiments with artificial neural networks, we generate a large-scale dataset of learning trajectories of aggregate statistics measured in a variety of neural network architectures, loss functions, learning rule hyperparameters, and parameter initializations. We then take a discriminative approach, training linear and simple non-linear classifiers to identify learning rules from features based on these observables.
Researcher Affiliation Academia 1Neurosciences Ph.D. Program, Stanford University 2Department of Computer Science, Stanford University 3Department of Applied Physics, Stanford University 4Department of Psychology, Stanford University 5Wu Tsai Neurosciences Institute, Stanford University
Pseudocode No The paper describes methods and procedures in narrative text and refers to figures, but it does not include any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes 3https://github.com/neuroailab/lr-identify
Open Datasets Yes We consider the tasks of supervised 1000-way Image Net categorization [Deng et al., 2009]... and supervised ten-way CIFAR-10 categorization [Krizhevsky, 2010].
Dataset Splits Yes test accuracy of each classifier, with mean and s.e.m. across ten category-balanced 75%/25% train/test splits, using the observable measures in 3.
Hardware Specification Yes We thank the Google Tensor Flow Research Cloud (TFRC) team for generously providing TPU hardware resources for this project.
Software Dependencies No The provided text does not specify any software names with version numbers (e.g., Python 3.8, PyTorch 1.9, TensorFlow 2.x) that would allow for replication of the experimental environment.
Experiment Setup Yes Learning hyperparameters for each model under a given learning rule category are the Cartesian product of three settings of batch size (128, 256, and 512)... All model training details can be found in Appendix A.