Generalization Bounds for Meta-Learning via PAC-Bayes and Uniform Stability

Authors: Alec Farid, Anirudha Majumdar

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

Reproducibility Variable Result LLM Response
Research Type Experimental We show that our bound provides a tighter guarantee than other bounds on a toy non-convex problem on the unit sphere and a text-based classification example. We also present a practical regularization scheme motivated by the bound in settings where the bound is loose and demonstrate improved performance over baseline techniques.
Researcher Affiliation Academia Alec Farid Anirudha Majumdar Department of Mechanical and Aerospace Engineering, Princeton University {afarid, ani.majumdar}@princeton.edu
Pseudocode Yes Algorithm 1 PAC-BUS: meta-learning via PAC-Bayes and Uniform Stability
Open Source Code Yes All the code required to run the following examples is available at https://github.com/irom-lab/PAC-BUS.
Open Datasets Yes Omniglot [35], Mini-Wiki benchmark introduced in [34]. This is derived from the Wiki3029 dataset presented in [9].
Dataset Splits Yes For all methods, 10,000 tasks are sampled as meta-training data and 1,000 tasks are sampled as held-out meta-test data. For the prior training step, all methods use 100 meta-training tasks.
Hardware Specification No It took multiple weeks of computation time on Amazon Web Services (AWS) instances to train and compute all networks and results we present in this paper. However, no specific instance types or hardware specifications (e.g., GPU/CPU models) are provided.
Software Dependencies Yes We use Python 3.8.3, learn2learn v0.1.0, PyTorch 1.5.0, torch-imle 0.0.1, and scipy 1.4.1.
Experiment Setup Yes We choose the softmax-activated cross-entropy loss, CELs, as the loss function. (from Section 5.1). Additionally, the Appendix A.11 provides tables of hyperparameters (e.g., 'Table 5: Hyperparameters for experiments on the classification on ball problem.' listing learning rates, epochs, etc.)