Task-level Differentially Private Meta Learning

Authors: Xinyu Zhou, Raef Bassily

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, we conduct several experiments demonstrating the effectiveness of our proposed algorithms.
Researcher Affiliation Academia Xinyu Zhou Department of Computer Science & Engineering The Ohio State University zhou.3542@buckeyemail.osu.edu Raef Bassily Department of Computer Science & Engineering and TDAI Institute The Ohio State University bassily.1@osu.edu
Pseudocode Yes Algorithm 1: Ameta-NSGD: meta learning with mini-batch noisy SGD
Open Source Code Yes Our code is available online at https://github.com/xyzhou055/Meta NSGD
Open Datasets Yes We consider linear regression task with mean square loss. Each task contains 10 datapoints (xi, yi)10 i=1. In Appendix C, we present additional experiments to evaluate our algorithms on Omniglot [28] few-shot classification tasks
Dataset Splits No The paper describes the number of data points per task ("Each task contains 10 datapoints") and the overall distribution of tasks, but does not explicitly provide specific train/validation/test splits for the collection of tasks used in the meta-learning experiments. The ethics checklist indicates this information might be present elsewhere, but it is not in the provided main text.
Hardware Specification No The paper explicitly states 'Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [No]'. No specific hardware details are mentioned in the main text.
Software Dependencies No The paper mentions using 'Tensor Flow privacy[27]' but does not specify its version number or any other software dependencies with their respective versions.
Experiment Setup Yes We let d = 30 for all experiments. Each task contains 10 datapoints (xi, yi)10 i=1. The privacy parameters ϵ are chosen from {1, 3, 10} and δ is set as 10 5. We set the clipping norm to 2. In the experiment, t, the number of cluster in the underlying distribution, is set to be 3, σ = 0.5 and we choose { h1, h2, h3} to be three orthogonal vectors with different norms.