Timezone: »

 
Poster
Overcoming Catastrophic Forgetting by Incremental Moment Matching
Sang-Woo Lee · Jin-Hwa Kim · Jaehyun Jun · Jung-Woo Ha · Byoung-Tak Zhang

Wed Dec 06 06:30 PM -- 10:30 PM (PST) @ Pacific Ballroom #37 #None
in Posters Wed »

Catastrophic forgetting is a problem of neural networks that loses the information of the first task after training the second task. Here, we propose a method, i.e. incremental moment matching (IMM), to resolve this problem. IMM incrementally matches the moment of the posterior distribution of the neural network which is trained on the first and the second task, respectively. To make the search space of posterior parameter smooth, the IMM procedure is complemented by various transfer learning techniques including weight transfer, L2-norm of the old and the new parameter, and a variant of dropout with the old parameter. We analyze our approach on a variety of datasets including the MNIST, CIFAR-10, Caltech-UCSD-Birds, and Lifelog datasets. The experimental results show that IMM achieves state-of-the-art performance by balancing the information between an old and a new network.

Keywords: Deep Learning · Bayesian Theory · Online Learning · Multitask and Transfer Learning

Author Information

Sang-Woo Lee (Naver Corp.)
Jin-Hwa Kim (SK T-Brain)

Research Scientist at SK T-Brain

Jaehyun Jun (Seoul National University)
Jung-Woo Ha (Clova AI Research, NAVER Corp.)
Byoung-Tak Zhang (Seoul National University & Surromind Robotics)

Related Events (a corresponding poster, oral, or spotlight)

More from the Same Authors