Loss Decoupling for Task-Agnostic Continual Learning

Authors: Yan-Shuo Liang, Wu-Jun Li

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments show that LODE can outperform existing state-of-the-art replay-based methods on multiple continual learning datasets.
Researcher Affiliation Academia Yan-Shuo Liang and Wu-Jun Li National Key Laboratory for Novel Software Technology, Department of Computer Science and Technology, Nanjing University, P. R. China liangys@smail.nju.edu.cn,liwujun@nju.edu.cn
Pseudocode Yes Algorithm 1 Loss Decoupling (LODE) for Continual Learning
Open Source Code No The paper states it's built on an existing repository ('mammoth [7] continual learning repository in Py Torch') but does not explicitly state that the code for their proposed method (LODE) is open-source or provided.
Open Datasets Yes Datasets We use three popular datasets for evaluation, including Seq-CIFAR10 [3], Seq CIFAR100 [10], and Seq-Tiny Image Net [22].
Dataset Splits Yes For each task, 5% of the training samples are divided into a validation set.
Hardware Specification No The paper does not specify the hardware used for experiments, such as particular GPU or CPU models.
Software Dependencies No The paper mentions using 'Py Torch' but does not provide its version number or specific versions for other software dependencies.
Experiment Setup Yes We use stochastic gradient descent (SGD) to optimize the parameters. The batch size and replay size are set to 32... We also follow existing methods [7, 5] to set memory as 500 and 5120 for all the datasets.