Meta-Learning Priors Using Unrolled Proximal Networks

Authors: Yilang Zhang, Georgios B. Giannakis

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Numerical tests conducted on few-shot learning datasets demonstrate markedly improved performance with flexible, visualizable, and understandable priors. and In this section, numerical tests are presented on several meta-learning benchmark datasets to evaluate the empirical performance of Meta Prox Net.
Researcher Affiliation Academia Yilang Zhang, Georgios B. Giannakis Department of Electric and Computer Engineering University of Minnesota Minneapolis, MN 55414, USA {zhan7453,georgios}@umn.edu
Pseudocode Yes Algorithm 1: Vanilla PGD algorithm for solving (1b) and Algorithm 2: Meta Prox Net algorithm
Open Source Code Yes All experiments are run on a server with RTX A5000 GPU, and our codes are available online at https://github.com/zhangyilang/Meta Prox Net.
Open Datasets Yes The mini Image Net dataset (Vinyals et al., 2016) consists of 60, 000 natural images sampled from the full Image Net (ILSVRC-12) dataset. and The Tiered Image Net (Ren et al., 2018) dataset is a larger subset of the Image Net dataset, composed of 779, 165 images from 608 classes.
Dataset Splits Yes The mini Image Net dataset... is split into 3 disjoint groups containing 64, 16 and 20 classes, which can be respectively accessed during the training, validation, and testing phases of meta-learning.
Hardware Specification Yes All experiments are run on a server with RTX A5000 GPU, and our codes are available online at https://github.com/zhangyilang/Meta Prox Net.
Software Dependencies No Adam optimizer is employed for tiered Image Net, while SGD with Nesterov momentum of 0.9 and weight decay of 10 4 is used for mini Image Net. The paper does not specify version numbers for any software dependencies.
Experiment Setup Yes The maximum number K of PGD steps (7) is 5, and the total number R of mini-batch SGD iterations (8) is 60, 000. The number of convolutional channels is 64 for Meta Prox Net+MAML, and 128 for Meta Prox Net+MC. The learning rates for PGD and SGD are α = 0.01 and β = 0.001, with batch size B = 4. Adam optimizer is employed for tiered Image Net, while SGD with Nesterov momentum of 0.9 and weight decay of 10 4 is used for mini Image Net. The interval [ A, A] and number C of pieces are determined through a grid search leveraging the validation tasks. For both mini Image Net and Tiered Imge Net datasets, A = 0.02 and C = 5.