Accurate Bayesian Meta-Learning by Accurate Task Posterior Inference

Authors: Michael Volpp, Philipp Dahlinger, Philipp Becker, Christian Daniel, Gerhard Neumann

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our empirical evaluation aims to study the effect on the predictive performance of (i) our improved TRNG-VI approach as well as of (ii) expressive variational GMM TP approximations in NP-based BML, in (iii) comparison to the state-of-the-art on (iv) a range of practically relevant meta-learning tasks.
Researcher Affiliation Collaboration 1Karlsruhe Institute of Technology, Karlsruhe, Germany 2Bosch Center for Artificial Intelligence, Renningen, Germany
Pseudocode Yes A.1.3 ALGORITHM SUMMARY We provide pseudocode for the meta-training stage of our GMM-NP algorithm in Alg. 1 and for the prediction stage in Alg. 2.
Open Source Code Yes We also provide source code for our proposed GMM-NP algorithm: Source code four our GMM-NP algorithm: https://github.com/ALRhub/gmm_np
Open Datasets Yes 2D MNIST Image Completion. We use the MNIST handwritten image database (Le Cun & Cortes, 2010).
Dataset Splits No The paper uses 'unseen test tasks' which implies a split, but does not explicitly provide percentages or counts for training/validation/test splits of the overall meta-datasets, nor does it define a separate validation set.
Hardware Specification No This work was performed on the Hore Ka supercomputer funded by the Ministry of Science, Research and the Arts Baden-W urttemberg and by the Federal Ministry of Education and Research. The authors further acknowledge support by the state of Baden-W urttemberg through bw HPC.
Software Dependencies No The paper cites TensorFlow and PyTorch as software used for automatic differentiation and mentions wandb for hyperparameter search, but does not provide specific version numbers for these or any other key software dependencies used in their experiments.
Experiment Setup Yes For a fair comparison, we employ a fixed experimental protocol for all datasets and models: we first perform a Bayesian hyperparameter search (HPO) to determine optimal algorithm settings, individually for each model-dataset combination. We then retrain the best model with 8 different random seeds and report the median log marginal predictive likelihood (LMLHD) as well as the median mean squared error (MSE), both in dependence of the context set size. (A.3.1 covers details such as DNN architectures, latent dimensionalities, number of GMM components, learning rates and trust region bounds with specific ranges and settings).