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

Xilai Li1*, Yingbo Zhou2*, Tianfu Wu1, Richard Socher2, and Caiming Xiong2 North Carolina State University1, Salesforce Research2

Task 1

Model 1

Task 2

Model 2

Task i-1

Model i-1

Task i

Model i

... ... Learn to Grow: A Continual Structure Learning Framework for Overcoming Catastrophic Forgetting Poster #89

• Fixed structure: Will

finally limited by the capacity

• Manually growing is

sub-optimal

S1 S2 S3 S4 TB TA TB TA S4 S4 , S3 S3 , S2 S2 ,

Input Reused weight New weight

S1 S1 ,

Fixed Structure (Regularization based method, e.g. EWC, Kirkpatrick et. al., 2016) Uniformly Growth for new task (Progressive Nets, Rusu et. al., 2016)

Learn to Grow: A Continual Structure Learning Framework for Overcoming Catastrophic Forgetting Poster #89

Input Reused weight New weight Adapter Task specific layer (prev) Task specific layer (current)

C1 C2 C3 C4 TB TA S3 S3

OR

+

Learn-to-Grow

Structure Search from options: “reuse”, “adaptation”, “new”

S1 S2 S3 S4 TB TA TB TA S4 S4 , S3 S3 , S2 S2 , S1 S1 ,

Fixed Structure (Regularization based method, e.g. EWC, Kirkpatrick et. al., 2016) Uniformly Growth for new task (Progressive Nets, Rusu et. al., 2016)

Learn to Grow: A Continual Structure Learning Framework for Overcoming Catastrophic Forgetting Poster #89

Dt

train

t-1 t t+1

Super Net Θt-1 Structure Optimization θt = f(Dt

train, Θt-1)

Parameter Optimization Transfer knowledge from previously learned tasks (Super Net Θt-1) Store new knowledge Θt= Θt-1⋃ θt Update “reused” knowledge

Learn to Grow: A Continual Structure Learning Framework for Overcoming Catastrophic Forgetting Poster #89

• The structure optimization results in “new”
• n the first layer and “reuse” for the rest.
• Ablations experiments validates the search

results.

Qualitative analysis on the Searched Structure on Task 2 (Task 1: ImageNet)

➢ Learned structure is sensible ○ Similar tasks tends to share more structure and parameters ○ 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