BNS: Building Network Structures Dynamically for Continual Learning

Authors: Qi Qin, Wenpeng Hu, Han Peng, Dongyan Zhao, Bing Liu

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on five datasets MNIST, CIFAR10, CIFAR-100, F-EMNIST and F-Celeba show that BNS markedly outperforms the existing state-of-the-art Task-CL baselines.
Researcher Affiliation Academia 1 Center for Data Science, AAIS, Peking University 2 Department of Information Science, School of Mathematical Sciences, Peking University 3 Wangxuan Institute of Computer Technology, Peking University 4 Department of Computer Science, University of Illinois at Chicago {qinqi, phan, wenpeng.hu, zhaody}@pku.edu.cn, liub@uic.edu
Pseudocode No The paper describes the BNS algorithm using text and figures, but does not provide structured pseudocode or algorithm blocks.
Open Source Code Yes The code of BNS can be found at: https://github.com/lalalaup6/BNS
Open Datasets Yes We evaluate the proposed BNS3 on five image classification datasets, three standard benchmarks MNIST [28], CIFAR10 and CIFAR-100 [26] and two additional datasets F-EMNIST and FCeleba [21]
Dataset Splits Yes We use ten percent of the training data of each task as the validation set for reward computation.
Hardware Specification Yes For all experiments, we use a single Nvidia RTX 2080Ti GPU.
Software Dependencies No In this paper, we use Res Net18 pre-trained on Image Net in Pytorch to calculate the task similarity.
Experiment Setup Yes For BNS, we use SGD as the optimizer with learning rate 0.1 to train the continual learner f( , θt) except F-EMNIST and F-Celeba (learning rate 0.01). The parameters of the agent (LSTM) is updated by the Adam optimizer using the learning rate 0.0001. The hyperparameters η and β are set to 0.001 and 0.003 respectively. To be consistent with the baseline settings, our continual learner trains 100 epochs for all datasets except MNIST (10 epochs).