Task Weighting in Meta-learning with Trajectory Optimisation

Authors: Cuong C. Nguyen, Thanh-Toan Do, Gustavo Carneiro

TMLR 2023 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental empirically demonstrate that the proposed approach out-performs common hand-engineering weighting methods on two few-shot learning benchmarks. (Abstract) 5 Experiments section begins on page 6, presenting performance metrics, figures, and tables such as "Figures 1a to 1d plot the testing accuracy evaluated on 100 validation tasks" and "Table 1: The classification accuracy averaged on 1,000 random testing tasks".
Researcher Affiliation Academia Cuong Nguyen EMAIL Australian Institute for Machine Learning The University of Adelaide, Thanh-Toan Do EMAIL Department of Data Science and AI Monash University, Gustavo Carneiro EMAIL Centre for Vision, Speech and Signal Processing University of Surrey. All listed affiliations are academic institutions.
Pseudocode Yes The complete algorithm of the proposed task-weighting meta-learning approach is outlined in Algorithm 1. (Page 6). Algorithm 1 Task-weighting for meta-learning (Page 6). Algorithm 2 Implementation of i LQR backward to determine the controller of interest. (Appendix D, Page 14).
Open Source Code No The paper does not contain any explicit statement about releasing code for the methodology described, nor does it provide a direct link to a code repository. Statements like "We release our code..." or links to GitHub are absent.
Open Datasets Yes The experiments are based on n-way k-shot classification setting used in few-shot learning with tasks formed from Omniglot (Lake et al., 2015) and mini-Image Net (Vinyals et al., 2016) the two most widely-used datasets in meta-learning.
Dataset Splits Yes For Omniglot, we follow the original train-test split (Lake et al., 2015) by using 30 alphabets for training and 20 alphabets for testing. For mini-Image Net, we follow the standard train-test split that uses 64 classes for training, 16 classes for validation and 20 for testing (Ravi & Larochelle, 2017) in our evaluation. The number of validation (or query) samples is kept at 15 samples per class.
Hardware Specification Yes Each experiment is carried out on a single NVIDIA Tesla V100 GPU with 32 GB memory following the configuration of NVIDIA DGX-1. The experiments mentioned in Section 5.4 is performed on a single NIVIDIA A100 GPU with 40 GB memory.
Software Dependencies No Appendix G, "Experimental settings," mentions that "GD is used to obtain ϕ (x)" and the "Adam optimiser" is used. However, it does not specify version numbers for any programming languages (e.g., Python), machine learning frameworks (e.g., PyTorch, TensorFlow), or other relevant software libraries.
Experiment Setup Yes For all experiments, GD is used to obtain ϕ (x) with 5 iterations and a learning rate of 0.1 for Omniglot and 0.01 for mini-Image Net. The learning rate for the meta-parameters, α, is set at 10 4 for all the setting. The mini-batch size is M = 10 tasks for Omniglot and M = 5 tasks for mini-Image Net. For the Dirichlet concentration of the prior in the exploration and exploitation baselines... we select κ = 1.2. The number of iterations used in i LQR is 2... The number of time steps (or number of mini-batches) is T = 10 for Omniglot and 5 for mini-Image Net. The parameters of the prior on the action ut are µu = 1/M and βu = 10.