Adversarial Continual Learning Sayna Ebrahimi Franziska Meier Roberto Calandra Trevor Darrell Marcus Rohrbach UC Berkeley Facebook AI Research Facebook AI Research UC Berkeley Facebook AI Research
What is Continual Learning? Definition: learning a sequence of tasks and performing well on all of them Objectives: Task 3 Task 2 Task 1 • No forgetting • Data streams and revisiting is not allowed/limited • High knowledge transferability • Efficiency and scalability • No/limited task information at test time • etc.
Approaches in Continual Learning LWF: Li & Hoiem, 2016 LWF: Li & Hoiem, 2016 LFL: Jung et al., 2016 LFL: Jung et al., 2016 EWC: Kirkpatrick et al., 2016 SI EWC EWC: Kirkpatrick et al., 2016 SI: Zenke et al., 2017 MAS SI: Zenke et al., 2017 VCL: Nguyen et al., 2017 MAS: Aljundi, 2018 MAS: Aljundi, 2018 Regularization UCB: Ebrahimi et al., 2020 UCB: Ebrahimi et al., 2020 LFL LWF UCB VCL
Approaches in Continual Learning LWF: Li & Hoiem, 2016 LWF: Li & Hoiem, 2016 LFL: Jung et al., 2016 LFL: Jung et al., 2016 EWC: Kirkpatrick et al., 2016 SI EWC EWC: Kirkpatrick et al., 2016 SI: Zenke et al., 2017 MAS SI: Zenke et al., 2017 VCL: Nguyen et al., 2017 MAS: Aljundi, 2018 MAS: Aljundi, 2018 Regularization UCB: Ebrahimi et al., 2020 UCB: Ebrahimi et al., 2020 LFL LWF UCB VCL A-GEM GEM Memory Experience replay : Robins, 1995 GEM: Lopez-Paz & Ranzato, 2017 Experience replay A-GEM: Chaudhry et al., 2019
Approaches in Continual Learning LWF: Li & Hoiem, 2016 LWF: Li & Hoiem, 2016 LFL: Jung et al., 2016 LFL: Jung et al., 2016 EWC: Kirkpatrick et al., 2016 SI EWC EWC: Kirkpatrick et al., 2016 SI: Zenke et al., 2017 MAS SI: Zenke et al., 2017 VCL: Nguyen et al., 2017 MAS: Aljundi, 2018 MAS: Aljundi, 2018 Regularization UCB: Ebrahimi et al., 2020 UCB: Ebrahimi et al., 2020 LFL LWF UCB VCL A-GEM PNN GEM Structure Memory Experience replay : Robins, 1995 PC GEM: Lopez-Paz & Ranzato, 2017 DEN Experience replay PNN : Rusu et al., 2016 A-GEM: Chaudhry et al., 2019 DEN : Yoon et al., 2018 PC: Schwarz et al., 2018
Approaches in Continual Learning LWF: Li & Hoiem, 2016 LWF: Li & Hoiem, 2016 LFL: Jung et al., 2016 LFL: Jung et al., 2016 EWC: Kirkpatrick et al., 2016 SI EWC EWC: Kirkpatrick et al., 2016 SI: Zenke et al., 2017 MAS SI: Zenke et al., 2017 VCL: Nguyen et al., 2017 MAS: Aljundi, 2018 MAS: Aljundi, 2018 Regularization UCB: Ebrahimi et al., 2020 UCB: Ebrahimi et al., 2020 LFL LWF UCB iCaRL: Rebuffi et al., 2016 VCL PackNet iCaRL Piggyback PackNet: Mallya, 2018 A-GEM Piggyback: Mallya 2018 PNN HAT HAT: Serrà et al., 2018 GEM DGR Structure Memory FearNet Experience replay : Robins, 1995 PC GEM: Lopez-Paz & Ranzato, 2017 DEN Experience replay PNN : Rusu et al., 2016 A-GEM: Chaudhry et al., 2019 DEN : Yoon et al., 2018 PC: Schwarz et al., 2018 DGR: Shin et al ., 2017 FearNet: Kemker et al., 2017
Where does our approach (ACL) stand? Memory Structure ACL
Intuition • Tasks in a sequence • Have task-invariant (shared) knowledge in common • Require task-specific (private) features to master them Private (Task 1) Shared Knowledge Private (Task 2) (task-invariant) Private (Task 3) How can we factorize task-invariant from task-specific features?
Our Approach: Adversarial Continual Learning Private (Task 1) Private Target label (Task 1) Private Private (Task 2) (Task 2) z P Private (Task 3) Private (Task 3) z S Shared Discriminator Task label Input images & task labels
Our Approach: Adversarial Continual Learning Private (Task 1) Private Target label (Task 1) Private Private (Task 2) (Task 2) z P Private (Task 3) Private (Task 3) z S Shared Discriminator Task label Input images & task labels
Our Approach: Adversarial Continual Learning Private (Task 1) Private Target label (Task 1) Private Private (Task 2) (Task 2) z P Private (Task 3) Private (Task 3) z S Shared Shared Discriminator Task label (task-invariant) Input images & task labels ℒ adv
Our Approach: Adversarial Continual Learning Private (Task 1) Private Target label (Task 1) Private Private (Task 2) (Task 2) z P Private (Task 3) Private (Task 3) z S Shared Shared Discriminator Task label (task-invariant) Input images & task labels ℒ adv
Our Approach: Adversarial Continual Learning Private (Task 1) Private Target label (Task 1) Private Private Target label (Task 2) (Task 2) z P Private (Task 3) Private (Task 3) ℒ di ff z S Shared Shared Shared Discriminator Task label (task-invariant) (task-invariant) Input images & task labels ℒ adv
Our Approach: Adversarial Continual Learning Private (Task 1) Private Target label (Task 1) Private Private Target label (Task 2) (Task 2) z P Private Target label (Task 3) Private (Task 3) ℒ di ff z S Shared Discriminator Task label (task-invariant) Input images & task labels (x,t) ℒ adv
Avoiding Forgetting in ACL Stored per task Private (Task 1) Private (Task 1) Private Private Target label (Task 2) (Task 2) z P Private (Task 3) Private (Task 3) ℒ di ff Experience z S Replay Shared Discriminator Task label (task-independent) Input images & task labels ℒ adv
Experiments miniImageNet (20 Tasks) Average Accuracy CIFAR100 (20 Tasks) n ACC = 1 ∑ R i , n Split MNIST (5 Tasks) n i =1 Permuted MNIST (10/20/30/40 Tasks) Backward Transfer: Sequence of 5 datasets: n 1 ∑ BWT = R i , n − R i , i (SVHN, CIFAR10, MNIST, FashionMNIST, NotMNIST) n − 1 i =1 Evaluation metrics Datasets
Results on 20-Split MiniImageNet ACC (%) ACL (Ours) 62.07 62.07 HAT 59.45 PNN 58.96 ER-RES 57.32 A-GEM 52.43 Ordinary Finetune 28.76 Accuracy (%) 0 63 70
Results on 20-Split MiniImageNet ACL (Ours) 62.07 62.07 0.00 HAT 59.45 -0.04 PNN 0.00 58.96 ER-RES 57.32 -11.34 A-GEM 52.43 -15.23 Ordinary Finetune 28.76 -64.23 -70 0 70 63
Results on 20-Split MiniImageNet ACL (Ours) 62.07 62.07 0.00 HAT 59.45 -0.04 PNN 58.96 0.00 ER-RES 57.32 -11.34 A-GEM 52.43 -15.23 Ordinary Finetune 28.76 -64.23 -70 0 70 63 Architecture Memory (MB) Replay Buffer (MB) ACL (Ours) 113.1 8.50 HAT 123.6 PNN 588.0 ER-RES 110.10 57.3 A-GEM 110.10 52.4 Ordinary Finetune 28.8 Memory (MB) 0 200 400 600
Results on Sequence of 5-Datasets (SVHN, CIFAR10, MNIST, FashionMNIST, NotMNIST) # Classes Training Test 50 212,785 48,365
Results on Sequence of 5-Datasets ACC (%) ACL (Ours) 78.55 78.55 UCB 76.34 Finetune 27.32 Accuracy (%) 0 80 (SVHN, CIFAR10, MNIST, FashionMNIST, NotMNIST) # Classes Training Test 79 50 212,785 48,365
Results on Sequence of 5-Datasets ACL (Ours) 78.55 78.55 -0.01 UCB 76.34 -1.34 Ordinary Finetune -42.12 27.32 -80 0 80 Architecture Memory (MB) ACL (Ours) 16.5 UCB 32.8 79 Finetune 16.5 0 13.333 26.667 40
Ablation Study on 20-Split miniImageNet ACC (%) BWT (%) Discriminator ℒ di ff Replay buffer X X X 62.07 0.00
Ablation Study on 20-Split miniImageNet ACC (%) BWT (%) Discriminator ℒ di ff Replay buffer - X X 52.07 -0.01 X X X 62.07 0.00
Ablation Study on 20-Split miniImageNet ACC (%) BWT (%) Discriminator ℒ di ff Replay buffer - X X 52.07 -0.01 X - X 57.66 -3.71 X X X 62.07 0.00
Ablation Study on 20-Split miniImageNet ACC (%) BWT (%) Discriminator ℒ di ff Replay buffer X X 52.07 -0.01 X X 57.66 -3.71 X X - 60.28 0.00 X X X 62.07 0.00
Visualizing Adversarial Learning Effect (20-Split miniImageNet) w. Discriminator w/o Discriminator Task Shared Private Shared Private Number Task 20 C1 C1 C1 C1 Without C2 C2 C2 C2 Dis C3 C3 C3 C3 C4 C4 C4 C4 C5 C5 C5 C5 T1 T1 T2 T2 T3 T3 T4 T4 T5 T5 Tasks 1-10 T6 T6 T7 T7 T8 T8 T9 T9 T10 T10
Visualizing Adversarial Learning Effect (20-Split miniImageNet) w. Discriminator w/o Discriminator Task Shared Private Shared Private Number Task 20 C1 C1 C1 C1 Without C2 C2 C2 C2 Dis C3 C3 C3 C3 C4 C4 C4 C4 C5 C5 C5 C5 T1 T1 T2 T2 T3 T3 T4 T4 T5 T5 Tasks 1-10 T6 T6 T7 T7 T8 T8 T9 T9 T10 T10
Visualizing Adversarial Learning Effect (20-Split miniImageNet) w. Discriminator w/o Discriminator Task Shared Private Shared Private Number Task 20 C1 C1 C1 C1 Without C2 C2 C2 C2 Dis C3 C3 C3 C3 C4 C4 C4 C4 C5 C5 C5 C5 T1 T1 T2 T2 T3 T3 T4 T4 T5 T5 Tasks 1-10 T6 T6 T7 T7 T8 T8 T9 T9 T10 T10
Visualizing Adversarial Learning Effect (20-Split miniImageNet) w. Discriminator w/o Discriminator Task Shared Private Shared Private Number Task 20 C1 C1 C1 C1 Without C2 C2 C2 C2 Dis C3 C3 C3 C3 C4 C4 C4 C4 C5 C5 C5 C5 T1 T1 T2 T2 T3 T3 T4 T4 T5 T5 Tasks 1-10 T6 T6 T7 T7 T8 T8 T9 T9 T10 T10
Visualizing Adversarial Learning Effect (20-Split miniImageNet) w. Discriminator w/o Discriminator Task Shared Private Shared Private Number Task 20 C1 C1 C1 C1 Without C2 C2 C2 C2 Dis C3 C3 C3 C3 C4 C4 C4 C4 C5 C5 C5 C5 T1 T1 T2 T2 T3 T3 T4 T4 T5 T5 Tasks 1-10 T6 T6 T7 T7 T8 T8 T9 T9 T10 T10
Recommend
More recommend