Revisiting Logistic-softmax Likelihood in Bayesian Meta-Learning for Few-Shot Classification

Authors: Tianjun Ke, Haoqun Cao, Zenan Ling, Feng Zhou

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

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we present the results for few-shot classification tasks including accuracy and uncertainty quantification. We consider two challenging standard benchmark datasets: the Caltech UCSD Birds [32] and mini-Image Net [24].
Researcher Affiliation Academia Center for Applied Statistics and School of Statistics, Renmin University of China School of EIC, Huazhong University of Science and Technology
Pseudocode Yes The complete training and test procedure is described in Alg. 1 of Appendix VII.
Open Source Code Yes Code is publicly available at https://github.com/keanson/revisit-logistic-softmax.
Open Datasets Yes We consider two challenging standard benchmark datasets: the Caltech UCSD Birds [32] and mini-Image Net [24].
Dataset Splits Yes Following the procedure of prior work [20], we employ a standard Conv4 architecture [31] as the backbone and assess our models under six different settings, including 1-shot and 5-shot scenarios for both in-domain and cross-domain tasks. ... we first tune the temperature parameter on the validation set for alignment and then compute the ECE and MCE on the test set.
Hardware Specification No The paper does not provide specific hardware details (e.g., exact GPU/CPU models, memory amounts, or detailed computer specifications) used for running its experiments.
Software Dependencies No The paper does not provide specific ancillary software details (e.g., library or solver names with version numbers) needed to replicate the experiment.
Experiment Setup Yes We train and evaluate our models (denoted as CDKT) with the ELBO loss (denoted as ML) using the default number of epochs from Patacchiola et al. [20], and with the predictive likelihood loss (denoted as PL) using 800 epochs. However, we find that the ML (τ < 1) domain-transfer experiment requires 800 epochs to avoid underfitting. We set the temperature parameter to τ = 0.5 for the 5-shot PL experiment of mini-Image Net and domain transfer, and τ = 0.2 for all other experiments.