Towards Sample-efficient Overparameterized Meta-learning

Authors: Yue Sun, Adhyyan Narang, Ibrahim Gulluk, Samet Oymak, Maryam Fazel

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

Reproducibility Variable Result LLM Response
Research Type Experimental Numerical experiments on real and synthetic data verify our insights on overparameterized meta-learning.
Researcher Affiliation Academia Yue Sun University of Washington yuesun@uw.edu Adhyyan Narang University of Washington adhyyan@uw.edu Halil Ibrahim Gulluk Bogazici University hibrahimgulluk@gmail.com Samet Oymak University of California, Riverside oymak@ece.ucr.edu Maryam Fazel University of Washington mfazel@uw.edu
Pseudocode Yes Algorithm 1 Constructing the optimal representation
Open Source Code Yes The code for this paper is in https://github.com/sunyue93/Rep-Learning.
Open Datasets Yes A Res Net-50 network pretrained on Imagenet was utilized to obtain a representation of R features for classification on CIFAR-10.
Dataset Splits Yes Data for Representation Learning (Phase 1). We have T tasks, each with n1 training examples... Data for Few-Shot Learning (Phase 2). Few-shot dataset has n2 examples (yi, xi)n2 j=1 i.i.d. Dβ. In Figure 1(a), 'Few shot train size = 20 Few shot train size = 100 Few shot train size = 500' are explicitly given.
Hardware Specification No The paper states '[N/A]' for question 3(d) regarding the total amount of compute and type of resources used for experiments, indicating no hardware specifications are provided.
Software Dependencies No The paper mentions software components like 'Res Net-50 network' but does not specify any software dependencies with version numbers.
Experiment Setup Yes A Res Net-50 network pretrained on Imagenet was utilized to obtain a representation of R features for classification on CIFAR-10. All layers except the final (softmax) layer are frozen and are treated as a fixed feature-map. We then train the final layer of the network for the downstream task which yields a linear classifier on pretrained features. In Fig. 5, we plot the error of the whole meta-learning algorithm. d = 100, n2 = 40, T = 200, ΣT = (I20, 0.05 I80), ΣF = I100.