Deciphering and Optimizing Multi-Task Learning: a Random Matrix Approach

Authors: Malik Tiomoko, Hafiz Tiomoko Ali, Romain Couillet

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

Reproducibility Variable Result LLM Response
Research Type Experimental This article provides theoretical insights into the inner workings of multi-task and transfer learning methods... Experiments on popular datasets demonstrate that our improved MTL LS-SVM method is computationally-efficient and outperforms sometimes much more elaborate state-of-the-art multi-task and transfer learning techniques.
Researcher Affiliation Collaboration Malik Tiomoko Laboratoire des Signaux et Systemes Universit e Paris-Sud Orsay, France... Hafiz Tiomoko Ali Huawei Technologies Research and Development (UK) London, UK... Romain Couillet Gipsa Lab Universit e Grenoble-Alpes Saint Martin d H eres, France
Pseudocode Yes Algorithm 1 Proposed Multi Task Learning algorithm.
Open Source Code Yes Reproducibility. Matlab and Julia codes for reproducing the results of the article are available in the supplementary materials.
Open Datasets Yes We next turn to the classical Office+Caltech256 (Saenko et al., 2010; Griffin et al., 2007) real data (images) benchmark for transfer learning, consisting of the 10 categories shared by both datasets.
Dataset Splits No The paper states 'Half of the samples of the target is randomly selected for the test data' but does not specify details for training or validation splits, nor mentions cross-validation.
Hardware Specification No The paper does not provide specific details on the hardware used for running experiments (e.g., GPU/CPU models, memory).
Software Dependencies No The paper mentions 'Matlab and Julia codes' are available in supplementary materials, but does not specify exact version numbers for these software or any other libraries/dependencies.
Experiment Setup Yes Figure 1 caption states: 'p = 100, [c11, c12, c21, c22] = [0.3, 0.4, 0.1, 0.2], γ = 12, λ = 10.' The paper also describes Algorithm 1 and other experimental settings in Section 5.