LegendreTron: Uprising Proper Multiclass Loss Learning

Authors: Kevin H Lam, Christian Walder, Spiridon Penev, Richard Nock

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

Reproducibility Variable Result LLM Response
Research Type Experimental Tested on a benchmark of domains with up to 1,000 classes, our experimental results show that our method consistently outperforms the natural multiclass baseline under a t-test at 99% significance on all datasets with greater than 10 classes.
Researcher Affiliation Collaboration 1School of Mathematics & Statistics, UNSW Sydney, Australia 2Google Research 3ANU College of Engineering, Computing and Cybernetics, The Australian National University, Australia 4UNSW Data Science Hub (u DASH), UNSW Sydney, Australia.
Pseudocode Yes Algorithm 1 describes LEGENDRETRON in detail.
Open Source Code No The paper does not contain an explicit statement about the release of the source code for the methodology described, nor does it provide a direct link to a code repository.
Open Datasets Yes For the three MNIST-like datasets (Le Cun et al., 2010; Xiao et al., 2017; Clanuwat et al., 2018)... We also compared LEGENDRETRON against multinomial logistic regression on 15 datasets that are publicly available from the LIBSVM library (Chang & Lin, 2011), the UCI machine learning repository (Asuncion & Newman, 2007; Dua & Graff, 2017), and the Statlog project (King et al., 1995).
Dataset Splits No The paper states: "each run randomly splits the dataset into 80% training and 20% testing sets." While it specifies training and testing splits, it does not explicitly mention a validation split or provide details for one.
Hardware Specification Yes Average GPU run times on a P100 for MNIST experiments in Table 2, were 2.32 and 2.12 hours for VGGTRON and VGG respectively.
Software Dependencies No The paper states: "All experiments were performed using Py Torch (Paszke et al., 2019)". While PyTorch is mentioned and cited, a specific version number for PyTorch itself, or other key software dependencies like Python or CUDA, is not provided.
Experiment Setup Yes For our experiments, we set softmax+ as the squashing function u for both LEGENDRETRON and multinomial logistic regression. We defer the full experimental details to Appendix I. Appendix I.1 includes a table "Network Architecture and Optimisation Details" specifying learning rate (α), weight decay (λ), multiplicative rate of decay (γ), epochs, and batch size for different datasets.