Overcoming Catastrophic Forgetting in Incremental Few-Shot Learning by Finding Flat Minima

Authors: Guangyuan SHI, JIAXIN CHEN, Wenlong Zhang, Li-Ming Zhan, Xiao-Ming Wu

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

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we empirically evaluate our proposed method for incremental few-shot learning and demonstrate its effectiveness by comparison with state-of-the-art methods.
Researcher Affiliation Academia Guangyuan Shi, Jiaxin Chen , Wenlong Zhang, Li-Ming Zhan, Xiao-Ming Wu Department of Computing The Hong Kong Polytechnic University {guang-yuan.shi, jiax.chen, wenlong.zhang, lmzhan.zhan}@connect.polyu.hk xiao-ming.wu@polyu.edu.hk
Pseudocode Yes Algorithm 1: F2M
Open Source Code Yes The source code is available at https://github.com/moukamisama/F2M.
Open Datasets Yes Datasets. For CIFAR-100 and mini Image Net, we randomly select 60 classes as the base classes and the remaining 40 classes as the new classes. ... For CUB-200-2011 with 200 classes, we select 100 classes as the base classes and 100 classes as the new ones.
Dataset Splits No The paper describes how training and test data are used but does not explicitly define a separate 'validation' dataset split for hyperparameter tuning. It states 'We tune the methods re-implemented by us to the best performance,' implying some validation was performed, but without specific split details.
Hardware Specification Yes The experiments are conducted with NVIDIA GPU RTX3090 on CUDA 11.0.
Software Dependencies Yes The experiments are conducted with NVIDIA GPU RTX3090 on CUDA 11.0.
Experiment Setup Yes In the base training stage, we select the last 4 or 8 convolution layers to inject noise... The flat region bound b is set as 0.01. We set the number of times for noise sampling as M = 2 4... In each incremental few-shot learning session, the total number of training epochs is 6, and the learning rate is 0.02.