Differentially Private Meta-Learning

Authors: Jeffrey Li, Mikhail Khodak, Sebastian Caldas, Ameet Talwalkar

ICLR 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, we apply our analysis to the problems of federated learning with personalization and few-shot classification, showing that allowing the relaxation to task-global privacy from the more commonly studied notion of local privacy leads to dramatically increased performance in recurrent neural language modeling and image classification.
Researcher Affiliation Collaboration Jeffrey Li, Mikhail Khodak, Sebastian Caldas Carnegie Mellon University jwl3@cs.cmu.edu Ameet Talwalkar Carnegie Mellon University & Determined AI
Pseudocode Yes Algorithm 1: Online version of our (ε, δ)-meta-private parameter-transfer algorithm.
Open Source Code No The paper mentions using 'tools provided by TensorFlow Privacy' and links to its GitHub, but does not provide a link or statement about releasing the source code for their specific methodology.
Open Datasets Yes Datasets: We train a LSTM-RNN for next word prediction on two federated datasets: (1) The Shakespeare dataset as preprocessed by (Caldas et al., 2018), and (2) a dataset constructed from 3, 000 Wikipedia articles drawn from the Wiki-3029 dataset (Arora et al., 2019), where each article is used as a different task. For few-shot image classification, we use the Omniglot (Lake et al., 2011) and Mini-Image Net (Ravi and Larochelle, 2017) datasets
Dataset Splits Yes For Shakespeare, we set the number of tokens per task to 800 tokens, leaving 279 tasks for meta-training, 31 for meta-validation, and 35 for meta-testing. For Wikipedia, we set the number of tokens to 1, 600, which corresponds to having 2, 179 tasks for meta-training, 243 for meta-validation, and 606 for meta-testing.
Hardware Specification No The paper does not provide specific hardware details such as GPU or CPU models, memory, or specific machine configurations used for running the experiments.
Software Dependencies No The paper mentions using 'TensorFlow Privacy' but does not specify its version number or any other software dependencies with version numbers.
Experiment Setup Yes Hyperparameters: For the language-modeling experiments, we tune the hyperparameters on the set of meta-validation tasks. For both datasets and all versions of the meta-learning algorithm, we tune hyperparameters in a two step process. We first tune all the parameters that are not related to refinement: the meta learning rate, the local (within-task) meta-training learning rate, the maximum gradient norm, and the decay constant. Then, we use the configuration with the best accuracy pre-refinement and then tune the refinement parameters: the refine learning rate, refine batch size, and refine epochs.