Dynamic Neural Network for Incremental Learning: Solution and Techniques
Addressing the challenge of incremental learning, this research presents a Dynamic Neural Network solution that enables training without previous data. The approach focuses on fast learning, reduced storage and memory costs, and optimal performance without forgetting past knowledge. Techniques such as network expansion, knowledge distillation, regularization, and rehearsal methods are employed to enhance learning across domains and tasks.
Download Presentation
Please find below an Image/Link to download the presentation.
The content on the website is provided AS IS for your information and personal use only. It may not be sold, licensed, or shared on other websites without obtaining consent from the author. Download presentation by click this link. If you encounter any issues during the download, it is possible that the publisher has removed the file from their server.
E N D
Presentation Transcript
Dynamic Neural Network for Incremental Learning Hikvision Research Institute Liang Ma, Jianwen Wu, Qiaoyong Zhong, Di Xie, Shiiliang PuQiaoyong Zhong, Di Xie, Shiiliang Pu
Content What we ask for incremental learning Main stream in the community Our solution: Dynamic Neural Network
What we ask for incremental learning Training without using previous data data- increme ntal Wheredata-incremental data&class- incremental class-incremental data&class- incremental class- increme ntal Training fast GOAL Less storage and memory costs Good performance without forgetting
Main stream in the community Regularization weight regularization: EWC/MAS/SI feature regularization: KD/SLNI Network Different networks for different tasks: PNN Rehearsal coreset selection and replay: iCarl/GEM generative replay
Our solution: Dynamic Neural Network Dynamic network expansion for data across dissimilar domain Knowledge distillation for data in similar domain Combined method of network expansion and feature regularization Features No need for previous data Dynamic network expansion to alleviate domain gap
Dynamic network expansion Freeze shared conv layers Network expansion for severe domain gap (bad accuracy) Tricks for generalization ability: For shared convs, imagenet pre-trained model For heads, more data augmentation and more batches to train head1
Knowledge distillation Replace BatchNorm with GroupNorm Mining known instances in new task and distill on best head known+ unknown known known shared convs shared convs shared convs head1 head1' head1 y ref_known y_known Cls loss KD loss
Experimental comparison Finetune DynamicNN
Experiment results in 1stround Square performance on all tasks Finetune 93.84 DynamicNN(No expand) 94.50 DynamicNN(expand@1) 95.75 DynamicNN(No expand) + LR trick 96.01 DynamicNN(expand@1) + LR trick 96.83 LR trick: Incremental learning with small learning rate