Toward Continual Learning on the Edge
0 likes•401 views
The document discusses continual learning, emphasizing its importance in improving AI scalability and adaptability, particularly for embedded systems and robotics. It outlines short-term research objectives, such as rehearsal-free, task-agnostic, real-time continual learning, and mentions the CORE50 dataset for continuous object recognition. Additionally, it covers future research directions focused on generative replay and enhancing real-world applications on devices like smartphones.
Toward Continual Learning on the Edge
- 1. Toward Continual Learning on the Edge Talk @ University of Pisa 14-02-2020 Vincenzo Lomonaco [email protected] Postdoctoral Researcher @ University of Bologna Supervisor: Davide Maltoni
- 2. About me • AI/Continual Learning Post-Doc @ University of Bologna • Co-Founder, President and Co-Director of Research at ContinualAI.org • Teaching Assistant of the courses Machine Learning and Computer Architectures @ UniBo • Author andTechnical reviewer of the online course Deep Learning with R and book R Deep Learning Essentials. • Co-Founder and Board Member of AIforPeople.org
- 3. What’s ContinualAI? • ContinualAI is a non-profit research organization and the largest research community on Continual Learning for AI. • It counts more than 600+ members in 17 different time-zones and from top-notch research institutions. • Learn more about ContinualAI at www.continualai.org
- 4. Machine Intelligence @ BioLab Davide Maltoni Vincenzo Lomonaco Lorenzo Pellegrini Gabriele Graffieti
- 5. Outline 1. Introduction and Motivation 2. Rehearsal-free and Task-agnostic Real-Time Continual Learning 3. Latent Replay for Real-Time Continual Learning 4. Real-world Deployment on Smartphone Devices
- 7. Machine Learning: State-of-the-art • Deep Learning holds state-of-the-art performances in many tasks. • Mainly supervised training with huge and fixed datasets.
- 8. The Curse of Dimensionality # of possible 227x227 RGB images
- 9. The Curse of Dimensionality # of possible 227x227 RGB images
- 10. The Curse of Dimensionality # of possible 227x227 RGB images
- 11. How can we improve AI scalability and adaptability? (Hence ubiquitousness and autonomy)
- 14. Continual Learning (CL) • Higher and realistic time-scale where data (and tasks) become available only during time. • No access to previously encountered data. • Constant computational and memory resources. • Incremental development of ever more complex knowledge and skills.
- 15. Practical Applications • Embedded systems and Robotics (+Privacy, +efficiency, +fast adaptation, +on the edge, -Internet connection) • AutoML and CI systems for AI models (+scalability, +efficiency, +fast adaptation, -energy consumption, -$$$) • Bias Removal (+Security patches, +fairness patches, +fast update)
- 16. The Stability-Plasticity Dilemma Stability-Plasticity Dilemma: • Remember past concepts • Learn new concepts • Generalize Biggest Problem in Deep Learning: • Catastrophic Forgetting
- 17. Continual Learning: Approaches T. Lesort,V. Lomonaco et al. Continual Learning for Robotics. pre-print arxiv arXiv:1907.00182 .
- 18. Rehearsal-free and Task-agnostic Real-Time Continual Learning
- 19. 3 Short-term Research Objectives for CL 1. Rehearsal-Free: Raw data cannot be stored and re-used for rehearsal. 2. Task Agnostic: No use of supplementary task supervised signal “t”. 3. Real-Time: Bounded computational and memory overheads, efficient, real-time updates. T. Lesort,V. Lomonaco et al. Continual Learning for Robotics. pre-print arxiv arXiv:1907.00182 .
- 20. Task Agnostic Continual Learning 1. New Instances (NI) 2. New Classes (NC) 3. New Instances and Classes (NIC) Initial Batch Incremental Batches Τ . . .
- 21. CORe50Website Dataset, Benchmark, code and additional information freely available at: vlomonaco.github.io/core50 LomonacoV. and Maltoni D. CORe50: a New Dataset and Benchmark for Continuous Object Recognition. CoRL2017.
- 22. LomonacoV. and Maltoni D. CORe50: a New Dataset and Benchmark for Continuous Object Recognition. CoRL2017. CORe50: aVideo Benchmark for CL and Object Recognition/Detection
- 23. # Images 164,866 Format RGB-D Image size 350x350 128x128 # Categories 10 # Obj. x Cat. 5 # Sessions 11 # img. x Sess. ~300 # Outdoor Sess. 3 Acquisition Sett. Hand held LomonacoV. and Maltoni D. CORe50: a New Dataset and Benchmark for Continuous Object Recognition. CoRL2017. CORe50: aVideo Benchmark for CL and Object Recognition/Detection
- 24. Fine-Grained Continual Learning LomonacoV., Maltoni D., Pellegrini L. Fine-Grained Continual Learning. Preprint arxiv arXiv:1907.03799, 2019.
- 25. AR-1* Rehearsal-free andTask Agnostic Real-Time Continual Learning Maltoni D. and LomonacoV. Continuous Learning in Single-Incremental-Task Scenarios. Neural Networks Journal, 2019. LomonacoV., Maltoni D., Pellegrini L. Fine-Grained Continual Learning. Preprint arxiv arXiv:1907.03799, 2019.
- 26. AR-1*: Overview (with MobileNet-V1) Data layer Output layer (classes) Low-level generic features Class specific discriminative features LomonacoV., Maltoni D., Pellegrini L. Fine-Grained Continual Learning. Preprint arxiv arXiv:1907.03799, 2019.
- 27. AR-1*: Supervised / Unsupervised Pre-Training Phase Data layer Output layer (classes) Low-level generic features Class specific discriminative features ● Supervised or Unsupervised Pre-Training from ImageNet. ● Slowly Fine-tuned or kept fixed. ● future direction: unsupervised co-training from scratch. LomonacoV., Maltoni D., Pellegrini L. Fine-Grained Continual Learning. Preprint arxiv arXiv:1907.03799, 2019.
- 28. AR-1*: Regularization Phase Data layer Output layer (classes) Low-level generic features Class specific discriminative features LomonacoV., Maltoni D., Pellegrini L. Fine-Grained Continual Learning. Preprint arxiv arXiv:1907.03799, 2019.
- 29. AR-1*: Regularization Phase Data layer Output layer (classes) Low-level generic features Class specific discriminative features ● Computational Efficient (independent from the number of training batches) ● Just one Fisher matrix (running sum + max clip) ● Importance of Batch ReNormalization LomonacoV., Maltoni D., Pellegrini L. Fine-Grained Continual Learning. Preprint arxiv arXiv:1907.03799, 2019.
- 30. AR-1*: Architectural Phase Data layer Output layer (classes) Low-level generic features Class specific discriminative features ● CWR*: generalization of CWR+ to handle agnostically NI, NC and NIC settings ● Dual-Memory system for memory consolidation. ● Based on zero-init for new classes, weights consolidation and finetuning for already encountered classes. LomonacoV., Maltoni D., Pellegrini L. Fine-Grained Continual Learning. Preprint arxiv arXiv:1907.03799, 2019.
- 31. CORe50 - NICv2 Results LomonacoV., Maltoni D., Pellegrini L. Fine-Grained Continual Learning. Preprint arxiv arXiv:1907.03799, 2019.
- 32. CORe50 - NICv2 Results LomonacoV., Maltoni D., Pellegrini L. Fine-Grained Continual Learning. Preprint arxiv arXiv:1907.03799, 2019. ● AR1* has a speed-up of x69 w.r.t. the Cumulative Strategy. ● AR1* is x393 more memory efficient than the Cumulative strategy. ● AR1* computation and memory overhead is fixed and independent from the number of training batches.
- 33. Latent Replay for Real-Time Continual Learning
- 34. AR-1*: Closing the Accuracy Gap with Latent Replay Data layer Output layer (classes) Low-level generic features Class specific discriminative features Backwardpass (allpatterns) Forwardpass (allpatterns) Forwardpass (nativepatterns) Concat External storage (rehearsal patterns) (at minibatch level) Pellegrini L., Graffieti G. , LomonacoV. and Maltoni D. Latent Replay for Real-Time Continual Learning.To be published. Backwardpass (nativepatterns)
- 35. AR-1*: Closing the Accuracy Gap with Latent Replay Pellegrini L., Graffieti G. , LomonacoV. and Maltoni D. Latent Replay for Real-Time Continual Learning.To be published.
- 36. AR-1*: Closing the Accuracy Gap with Latent Replay Pellegrini L., Graffieti G. , LomonacoV. and Maltoni D. Latent Replay for Real-Time Continual Learning.To be published.
- 37. AR-1*: Sparse Representations ● Imposing Sparsity of the activactivation does not affect accuracy from ~55% to ~35%. ● It has been shown that sparsity may help the CL process. ● Less memory overhead for latent replay. Pellegrini L., Graffieti G. , LomonacoV. and Maltoni D. Latent Replay for Real-Time Continual Learning.To be published.
- 38. Real-world Deployment on Smartphone Devices
- 39. CORe Android App ● AR1*free with Latent Replay (RM=500). ● 20 sec to gather new session images (at 5 fps). ● Near real-time training updates CPUs-only (about 1 second). ● Inference time of about 5fps without any specific optimization. Pellegrini L., Graffieti G. , LomonacoV. and Maltoni D. Latent Replay for Real-Time Continual Learning.To be published.
- 40. CORe Android App Pellegrini L., Graffieti G. , LomonacoV. and Maltoni D. Latent Replay for Real-Time Continual Learning.To be published.
- 41. FutureWorks and Research Direction 1. Latent Generative Replay 2. Lowering the amount of Supervision (Unsupervised Reinforcement Learning, Active Learning) 3. Infer or make use of the sparse “task signal” (context modulation) 4. Sequence Learning/ Temporal Coherence Integration 5. Improve robustness in real-world embedded applications (Smartphone devices, Nao Robot, …) Maltoni D. and LomonacoV. Semi-SupervisedTuning fromTemporal Coherence. ICPR 2016. LomonacoV., Desai K., Maltoni D. and Culurciello, E. Continual Reinforcement Learning in 3D non-stationary environments. preprint arxiv arXiv:1905.10112, 2019.
- 42. AR-1*: Closing the Accuracy Gap with Latent Generative Replay ● Two levels (category + classes) ● Dense connections, two losses ● Supervised pre-training on a large dataset (e.g. Imagenet). ● Unsupervised slow tuning based on i) temporal coherence or ii) reconstruction of input. ● Supervised training with latent rehearsal ● Regularization by temporal coherence (category level?) Generative Model Pellegrini L., Graffieti G. , LomonacoV. and Maltoni D. Latent Replay for Real-Time Continual Learning.To be published. Backwardpass (allpatterns) Forwardpass (allpatterns) Forwardpass (nativepatterns) Concat (at minibatch level) Backwardpass (nativepatterns)
- 43. Questions? Talk @ University of Pisa 14-02-2020 Vincenzo Lomonaco [email protected] Postdoctoral Researcher @ University of Bologna Supervisor: Davide Maltoni
- 44. Questions? Talk @ University of Pisa 14-02-2020 Vincenzo Lomonaco [email protected] Postdoctoral Researcher @ University of Bologna Supervisor: Davide Maltoni