To preserve the accuracy of the deployed models, we often need to continuously update the models. However, naive fine-tuning methods are effective and efficient due to forgetting, overfitting, and high resource demanding issues. Thus, we disucss the potential solutions here from algorithm and system perspectives.

alt text

Table of contents

Algorithms

  • [ICML 2017] Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks​
  • [NeurIPS’18] Memory Replay GANs: Learning to Generate New Categories without Forgetting​
  • [ICLR’20] Continual learning with hypernetworks​
  • [ICLR’22] Model Zoo: A Growing Brain That Learns Continually​
  • [ICML’22] Forget-free Continual Learning with Winning Subnetworks​
  • [NeurIPS’22] SparCL: Sparse Continual Learning on the Edge​
  • [CVPR 2022] Meta-attention for ViT-backed Continual Learning​
  • [ICML’23] Lifelong Language Pretraining with Distribution-Specialized Experts​
  • [CVPR’23] CODA-Prompt: Continual Decomposed Attention-based Prompting for Rehearsal-Free Continual Learning​
  • [CVPR’23] Task Difficulty Aware Parameter Allocation & Regularization for Lifelong Learning​
  • [ICCV’23] Self-Evolved Dynamic Expansion Model for Task-Free Continual Learning​
  • [ICCV’23] Label-Efficient Online Continual Object Detection in Streaming Video​
  • [ICLR’24] Divide and not forget: Ensemble of selectively trained experts in Continual Learning​
  • [ICLR’24] Continual Learning on a Diet: Learning from Sparsely Labeled Streams Under Constrained Computation ​
  • [ICLR 2024] TiC-CLIP: Continual Training of CLIP Models ​
  • [ICLR 2024] Scalable Language Model with Generalized Continual Learning ​
  • [ICLR 2024] A Unified and General Framework for Continual Learning​
  • [ICLR 2024] Divide and not forget: Ensemble of selectively trained experts in Continual Learning​
  • [AAAI’24] Summarizing Stream Data for Memory-Restricted Online Continual Learning​
  • [AAAI’24] Evolving Parameterized Prompt Memory for Continual Learning​

Syetems

  1. [ATC’22] SOTER: Guarding Black-box Inference for General Neural Networks at the Edge​
  2. [DAC’22] CarM: hierarchical episodic memory for continual learning​
  3. [DAC’22] The Larger The Fairer? Small Neural Networks Can Achieve Fairness for Edge Devices​
  4. [NSDI’22] Ekya: Continuous Learning of Video Analytics Models on Edge Compute Servers​
  5. [SOSP’23] Sia: Heterogeneity-aware, goodput-optimized ML-cluster scheduling​
  6. [ATC’23] EnvPipe: Performance-preserving DNN Training Framework for Saving Energy​
  7. [ATC’23] Beware of Fragmentation: Scheduling GPU-Sharing Workloads with Fragmentation Gradient Descent​
  8. [MobiCom’23] AccuMO: Accuracy-Centric Multitask Offloading in Edge-Assisted Mobile Augmented Reality, Purdue University​
  9. [EuroSys’23] Egeria: Efficient DNN Training with Knowledge-Guided Layer Freezing​
  10. [MOBICOM’23] Cost-effective On-device Continual Learning over Memory Hierarchy with Miro​
  11. [VLDB’23] RALF: Accuracy-Aware Scheduling for Feature Store Maintenance​
  12. [VLDB’23] Cache Me If You Can: Accuracy-Aware Inference Engine for Differentially Private Data Exploration​
  13. [NSDI’23] ModelKeeper: Accelerating DNN Training via Automated Training Warmup​
  14. [NSDI’23] RECL: Responsive Resource-Efficient Continuous Learning for Video Analytics​
  15. [DAC’23] Shoggoth: Towards Efficient Edge-Cloud Collaborative Real-Time Video Inference via Adaptive Online Learning​
  16. [EuroSys’24] Model Selection for Latency-Critical Inference Serving​
  17. [OSDI’24] Caravan: Practical Online Learning of In-Network ML Models with Labeling Agents ​
  18. [VLDB’24] Accelerating String-key Learned Index Structures via Memoization-based Incremental Training​
  19. [ISCA’24] DACAPO: Accelerating Continuous Learning in Autonomous Systems for Video Analytics​
  20. [DAC’24] Enabling On-Device Large Language Model Personalization with Self-Supervised Data Selection and Synthesis​
  21. [OSDI’24] Automatic and Efficient Customization of Neural Networks for ML Applications​
  22. [NSDI’24] MadEye: Boosting Live Video Analytics Accuracy with Adaptive Camera Configurations​
  23. [ASPLOS’24] Proteus: A High-Throughput Inference-Serving System with Accuracy Scaling​
  24. [ASPLOS’24] RAP: Resource-aware Automated GPU Sharing for Multi-GPU Recommendation Model Training and Input Preprocessing​
  25. [ASPLOS’24] NeuPIMs: NPU-PIM Heterogeneous Acceleration for Batched LLM Inferencing​
  26. [ASPLOS’24] SpecInfer: Accelerating Large Language Model Serving with Tree-based Speculative Inference and Verification​
  27. [arXiv 2024.01] MoE-Infinity: Activation-Aware Expert Offloading for Efficient MoE Serving​
  28. [arXiv 2024.03] Are More LLM Calls All You Need? Towards Scaling Laws of Compound Inference Systems
  29. [arXiv 2023.11] S-LoRA: Serving Thousands of Concurrent LoRA Adapters ​
  30. [arXiv 2023.10] Punica: Multi-Tenant LoRA Serving​
  31. [arXiv 2024.01] CARASERVE: CPU-Assisted and Rank-Aware LoRA Serving for Generative LLM Inference​
  32. [arXiv 2024.02] APIServe: Efficient API Support for Large-Language Model Inferencing​
  33. [arXiv 2024.02] FlexLLM: A System for Co-Serving Large Language Model Inference and Parameter-Efficient Finetuning​
  34. [arXiv 2024.03] ALTO: An Efficient Network Orchestrator for Compound AI Systems​
  35. [arXiv 2024.04] Mélange: Cost Efficient Large Language Model Serving by Exploiting GPU Heterogeneity​

​

Conclusion

References

​