Equivariance with Learned Canonicalization Functions

Authors: Sékou-Oumar Kaba, Arnab Kumar Mondal, Yan Zhang, Yoshua Bengio, Siamak Ravanbakhsh

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our main hypothesis, supported by our empirical results, is that learning a small neural network to perform canonicalization is better than using predefined heuristics. Our experiments show that learning the canonicalization function is competitive with existing techniques for learning equivariant functions across many tasks, including image classification, N-body dynamics prediction, point cloud classification and part segmentation, while being faster across the board.
Researcher Affiliation Collaboration 1School of Computer Science, Mc Gill University, Montr eal, Canada 2Mila Quebec Artficial Intelligence Institute, Montr eal, Canada 3Samsung SAIT AI Lab, Montr eal, Canada 4DIRO, Universit e de Montr eal, Montr eal, Canada.
Pseudocode Yes I Algorithm for Image Inputs Algorithm 1 Differentiable Canonicalization for Image Inputs
Open Source Code Yes Our code is available at: https://github.com/ oumarkaba/canonical_network.
Open Datasets Yes We selected the Rotated MNIST dataset (Larochelle et al., 2007), often used as a benchmark for equivariant architectures. [...] We use the Model Net40 (Wu et al., 2015) and Shape Net (Chang et al., 2015) datasets for experiments on point clouds.
Dataset Splits Yes We perform early stopping based on the classification performance of the validation dataset with a patience of 20 epochs.
Hardware Specification Yes Inference time (in seconds) of the networks for Model Net40 classification test split in 1 A100 and 8 CPUs with a batch size of 32.
Software Dependencies No The paper mentions software like Adam, batch-norm, ReLU, Kornia, and a PYTORCH code snippet, but does not provide specific version numbers for any of these software components or libraries.
Experiment Setup Yes In all our image experiments, we train the models by minimizing the cross entropy loss for 100 epochs using Adam (Kingma & Ba, 2014) with a learning rate of 0.001. We perform early stopping based on the classification performance of the validation dataset with a patience of 20 epochs.