Unconstrained Stochastic CCA: Unifying Multiview and Self-Supervised Learning

Authors: James Chapman, Lennie Wells, Ana Lawry Aguila

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our algorithms show far faster convergence and recover higher correlations than the previous state-of-the-art on all standard CCA and Deep CCA benchmarks.
Researcher Affiliation Academia James Chapman*, Ana Lawry Aguila University College London {james.chapman.19, ana.aguila.18}@ucl.ac.uk Lennie Wells* University of Cambridge ww347@cam.ac.uk
Pseudocode Yes Algorithm 2: DMCCA-EY Loss Function in Python
Open Source Code Yes To ensure the reproducibility of our work, we have made the following resources publicly available: Code for stochastic CCA and deep CCA experiments: https://github.com/jameschapman19/GEP-EY. For Self-Supervised Learning experiments, our modified version of solo-learn is accessible at: https://github.com/jameschapman19/solo-learn. Standalone Py Torch implementation of the SSL-EY loss function: https://github.com/jameschapman19/SSL-EY.
Open Datasets Yes We apply stochastic Partial Least Squares (PLS) to an extremely high-dimensional dataset from the UK Biobank (Sudlow et al., 2015)
Dataset Splits No The paper mentions evaluating on a 'validation set' for various experiments (e.g., 'empirical correlations between the representations on a validation set', 'Top-1 and Top-5 accuracies on the validation set') but does not provide specific details on how this validation set is split or its size/proportion for any dataset.
Hardware Specification Yes Experiment CPU/GPU Resources Stochastic CCA NVIDIA Ge Force RTX 2080 Ti Deep CCA NVIDIA Ge Force RTX 2080 Ti Deep MCCA NVIDIA Ge Force RTX 2080 Ti Stochastic PLS NVIDIA Ge Force GTX 1650 Ti SSL 4-8 NVIDIA Ge Force RTX 2080 Ti, Quadro RTX 8000 Quadro RTX 6000, or NVIDIA Ge Force GTX 1080 Ti GPU devices
Software Dependencies No The paper mentions using 'solo-learn' and 'PyTorch' (via pseudocode) but does not provide specific version numbers for these or other software dependencies.
Experiment Setup Yes Parameters: For each method, we searched over a hyperparameter grid using Biewald (2020). Parameter Values minibatch size 5,20,50,100 components 5 epochs 1 seed 1, 2, 3, 4, 5 lr 0.01, 0.001, 0.0001 γ10 0.01,0.1,1,10