Model Sensitivity Aware Continual Learning

Authors: Zhenyi Wang, Heng Huang

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments conducted across multiple datasets demonstrate the efficacy and versatility of our proposed method.
Researcher Affiliation Academia Zhenyi Wang and Heng Huang Department of Computer Science, Institute of Health Computing University of Maryland College Park College Park, MD, 20742 zwang@umd.edu;heng@umd.edu
Pseudocode Yes Algorithm 1 Model Sensitivity Aware Continual Learning
Open Source Code No Justification: Code will be released.
Open Datasets Yes We conduct experiments on several datasets, including CIFAR10 (10 classes), CIFAR100 (100 classes) [29], and Tiny-Image Net (200 classes) [80], to assess the effectiveness of MACL in task incremental learning (Task-IL) and class incremental learning (Class-IL).
Dataset Splits No Following [7, 14], the hyperparameter is determined through the validation sets split from the training sets from the first three tasks.
Hardware Specification Yes We use a single NVIDIA A5000 GPU with 24GB memory to run the experiments.
Software Dependencies No The paper mentions using 'standard SGD optimizer' but does not specify versions for programming languages, libraries (e.g., PyTorch, TensorFlow), or other software dependencies.
Experiment Setup Yes For the hyperparameters in our method, we set α = 1.0 across all the datasets to minimize the model’s dependence on hyperparameters. For η, we set η = 1e 5 for CIFAR10 and CIFAR100, and η = 1e 6 for Tiny-Image Net. The η is selected from the range of [1e 4, 1e 5, 1e 6, 1e 7]. ... The batch size and replay buffer batch size are set to 32.