Haizhou Shi Hao Wang Computer Science Department Rutgers University 101723 1 Background Domain Incremental Learning DIL Machine learning models ID: 1045787
Download Presentation The PPT/PDF document "A Unified Approach to Domain Incremental..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
1. A Unified Approach to Domain Incremental Learning with Memory: Theory and AlgorithmHaizhou Shi, Hao WangComputer Science Department, Rutgers University10/17/231
2. BackgroundDomain Incremental Learning (DIL)Machine learning models are required to incrementally learn the evolving data distributions.E.g., autonomous driving under different weather conditions.Memory constraint: no (or very limited size of) the past data can be stored during training. 10/17/232Continual Learner ... Current Domain LearningPrevious Domains(Unavailable)Continual Learner ... Current Domain LearningMemorySample SelectionSample Replay
3. BackgroundDomain Incremental Learning (DIL)Machine learning models are required to incrementally learn the evolving data distributions.E.g., autonomous driving under different weather conditions.Memory constraint: no (or very limited size of) the past data can be stored during training. Goal of DIL: minimize the model’s risk over all domains.10/17/233
4. ERM-Based Generalization BoundEmpirical Risk Minimization (ERM) via Experience Replay (ER)[Lemma 3.1] Trivially replaying the memory will cause a loose generalization bound.10/17/234Encoder Predictor Memory: Current Domain Data: Domain IDCurrent Model Classification Loss
5. Intra-Domain Model-Based BoundDark Experience Replay (DER++)[Lemma 3.2] Intra-Domain Model-Based Bound10/17/235Encoder Predictor History Model Memory: Current Domain Data: Domain IDCurrent Model Classification Loss Distillation Loss
6. Cross-Domain Model-Based BoundLearning without Forgetting (LwF)[Lemma 3.3] Cross-Domain Model-Based Bound 10/17/236Encoder Predictor History Model Current Domain Data: Domain IDCurrent Model Classification Loss Distillation Loss
7. UDIL: A Unified Bound for DILA set of coefficients integrates them into one unified bound.[Theorem 3.4] Unified Generalization Bound for all domains10/17/237Naïve ERMIntra-Domain BoundCross-Domain Bound
8. UDIL: A Unified Bound for DILUDIL unifies multiple existing methods under certain conditions.10/17/238
9. UDIL: An Adaptive Bound for DILUDIL can adaptively adjust the coefficients based on the data and the history model .It will, ideally, minimize the tightest bound in the family of all the generalization bounds. 10/17/239Encoder Predictor Discriminator History Model Domain ID, ,
10. UDIL: An Adaptive Bound for DIL10/17/2310Encoder Predictor Discriminator History Model Domain IDCross-Entropy Classification Loss
11. UDIL: An Adaptive Bound for DIL10/17/2311Encoder Predictor Discriminator History Model Domain IDCross-Entropy Distillation Loss
12. UDIL: An Adaptive Bound for DIL10/17/2312Encoder Predictor Discriminator History Model Domain IDAdversarial Feature Alignment Loss
13. UDIL: An Adaptive Bound for DIL10/17/2313Encoder Predictor Discriminator History Model Domain ID Adaptive Coefficient Optimization
14. UDIL: Experimental ResultsUDIL’s representation distribution on synthetic dataset (high-dimensional balls)10/17/2314
15. UDIL: Experimental ResultsUDIL evaluated on realistic datasets.10/17/2315HD-Balls, Permuted-MNIST, Rotated-MNIST
16. UDIL: Experimental ResultsUDIL evaluated on realistic datasets.10/17/2316Sequential CORe-50
17. ConclusionProposed a principled framework, UDIL, for domain incremental learning with memory to unify various existing methods. Theoretical analysis shows that different existing methods are equivalent to minimizing the same error bound with different fixed coefficients. UDIL allows adaptive coefficients during training, thereby always achieving the tightest bound and improving the performance. 10/17/2317