Towards Gradient-based Bilevel Optimization with Non-convex Followers and Beyond

Authors: Risheng Liu, Yaohua Liu, Shangzhi Zeng, Jin Zhang

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

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we first verify the theoretical convergence results on non-convex numerical problems compared with existing EG methods and IG methods. Then we test the performance of IAPTT-GM and demonstrate its generalizability to real-world BLO problems with non-convex followers, which are caused by non-convex regularization and neural network structures. In addition, we further validate the performance of the accelerated version (i.e., IA-GM (A)) under LLC with numerical examples and data hyper-cleaning tasks.
Researcher Affiliation Collaboration Risheng Liu1,2 Yaohua Liu1 Shangzhi Zeng3 Jin Zhang 4,5 1International School of Information Science & Engineering, DUT 2Pazhou Lab, Guangzhou 3Department of Mathematics and Statistics, UVic 4Department of Mathematics, SUSTech 5National Center for Applied Mathematics Shenzhen
Pseudocode Yes Algorithm 1 The Proposed IAPTT-GM
Open Source Code Yes The code is available at http://github.com/vis-opt-group/IAPTT-GM.
Open Datasets Yes We report results of IAPTT-GM and various mainstream methods, e.g., Prototypical Network [48], Relation Net [49] and T-RHG [16] on mini Image Net [47] and tiered Image Net [50] datasets with two different backbones [6, 51] in Table 2. and IAPTT-GM achieves better test performance of both accuracy and F1 score on two datasets, including MNIST [46] and Fashion MNIST [52].
Dataset Splits Yes the dataset is randomly split to three disjoint subsets: Dtr for training, Dval for validation and Dtest for testing and Information about the dataset and hyperparameters for numerical experiments can be found in the supplemental material.
Hardware Specification No The paper states: 'We specify the detailed configuration of computing devices, GPUs in the supplementary materials.', indicating that hardware specifications are not provided in the main text.
Software Dependencies No The main paper does not explicitly state software dependencies with specific version numbers. Details regarding experimental setup are generally referred to the supplementary materials.
Experiment Setup No The paper states that 'Information about the dataset and hyperparameters for numerical experiments can be found in the supplemental material.', indicating that specific experimental setup details are not provided in the main text.