Xilai Li 1* , Yingbo Zhou 2* , Tianfu Wu 1 , Richard Socher 2 , and - PowerPoint PPT Presentation

Xilai Li 1* , Yingbo Zhou 2* , Tianfu Wu 1 , Richard Socher 2 , and Caiming Xiong 2 North Carolina State University 1 , Salesforce Research 2

Task 1 Task 2 Task i-1 Task i ... ... Model 1 Model 2 Model i-1 Model i Learn to Grow: A Continual Structure Learning Framework for Overcoming Catastrophic Forgetting Poster #89

, S 1 S 1 S 1 , S 2 S 2 S 2 ● Fixed structure: Will , finally limited by the S 3 S 3 S 3 capacity , S 4 S 4 S 4 ● Manually growing is T A T B T A T B sub-optimal Uniformly Growth for new Fixed Structure ( Regularization task (Progressive Nets, based method , e.g. EWC, Rusu et. al., 2016) Kirkpatrick et. al., 2016) Input Reused weight New weight Learn to Grow: A Continual Structure Learning Framework for Overcoming Catastrophic Forgetting Poster #89

, S 1 S 1 S 1 C 1 , S 3 S 3 S 2 S 2 S 2 C 2 + , S 3 S 3 S 3 C 3 OR , S 4 S 4 S 4 C 4 Structure Search from options: “reuse”, “adaptation”, “new” T A T B T A T B T A T B Uniformly Growth for new Fixed Structure ( Regularization Learn-to-Grow task (Progressive Nets, based method , e.g. EWC, Rusu et. al., 2016) Kirkpatrick et. al., 2016) Input Reused weight New weight Adapter Task specific layer (prev) Task specific layer (current) Learn to Grow: A Continual Structure Learning Framework for Overcoming Catastrophic Forgetting Poster #89

t-1 t t+1 D t train Update “reused” Transfer knowledge from Structure knowledge previously learned tasks Optimization Parameter (Super Net Θ t-1 ) Optimization train , Θ t-1 ) θ t = f(D t Store new knowledge Θ t = Θ t-1 ⋃ θ t Super Net Θ t-1 Learn to Grow: A Continual Structure Learning Framework for Overcoming Catastrophic Forgetting Poster #89

Qualitative analysis on the Searched Structure on Task 2 (Task 1: ImageNet) ● The structure optimization results in “ new ” on the first layer and “reuse” for the rest. Learned structure is sensible ➢ ○ Similar tasks tends to share more ● Ablations experiments validates the search structure and parameters results. ○ Distant tasks share less Learn to Grow: A Continual Structure Learning Framework for Overcoming Catastrophic Forgetting Poster #89

Comparison between “ tune reuse ” and “ fix reuse ” ● The “tune” higher than “fix” at certain task indicates “ positive forward transfer ” ● The “tune” curve “goes up” means “ positive backward transfer ” Learn to Grow: A Continual Structure Learning Framework for Overcoming Catastrophic Forgetting Poster #89

Permuted-MNIST Split-CIFAR100 Learn to Grow: A Continual Structure Learning Framework for Overcoming Catastrophic Forgetting Poster #89