Neural Mechanics: Symmetry and Broken Conservation Laws in Deep Learning Dynamics

Authors: Daniel Kunin, Javier Sagastuy-Brena, Surya Ganguli, Daniel LK Yamins, Hidenori Tanaka

ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically validate our analytic expressions for learning dynamics on VGG-16 trained on Tiny Image Net.
Researcher Affiliation Collaboration Stanford University Physics & Informatics Laboratories, NTT Research, Inc.
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes An open source version of our code, used to generate all the figures in this paper, is available at github.com/danielkunin/neural-mechanics.
Open Datasets Yes Dataset. While we ran some initial experiments on Cifar-100, the dataset used in all the empirical figures in this documents was Tiny Imagenet.
Dataset Splits No The paper mentions 'Tiny Image Net' and training parameters like '100 epochs' and 'batch size S = 256', but does not specify explicit training/validation/test split percentages or sample counts, nor does it refer to predefined splits with citations.
Hardware Specification No The paper does not explicitly describe the hardware used for experiments, such as specific GPU/CPU models or detailed computer specifications.
Software Dependencies No The paper mentions 'Py Torch' as a tool but does not provide specific version numbers for any software dependencies.
Experiment Setup Yes All models were initialized using Kaiming Normal, and no learning rate drops or warmup were used. Model Dataset Epochs Batch size Opt. LR Mom. WD Damp. VGG-16 Tiny Image Net 100 256 SGD [0.1, 0.01] [0, 0.001, 0.0005, 0.0001] 0 VGG-16 w/BN Tiny Image Net 100 256 SGD [0.1, 0.01] [0, 0.001, 0.0005, 0.0001] 0 VGG-16 Tiny Image Net 100 128 SGDM 0.1 [0, 0.9, 0.99] [0, 0.001, 0.0005, 0.0001] 0 VGG-16 w/BN Tiny Image Net 100 128 SGDM 0.1 [0, 0.9, 0.99] [0, 0.001, 0.0005, 0.0001] 0