A Major Project by Abhinav
Continual learning systems struggle with catastrophic forgetting—models lose performance on old tasks when exposed to new ones. Rehearsal methods such as iCaRL replay stored exemplars to retain knowledge, yet they do so at a constant rate, wasting computation when the model is already stable. Meanwhile, the data-stream mining community has mature concept drift detectors (e.g., ADWIN) that flag statistically significant performance drops.
Core Hypothesis: By combining iCaRL with ADWIN we can build an adaptive rehearsal strategy that reacts only when forgetting is detected, preserving accuracy while reducing rehearsal cost.
- Adaptive rehearsal policy: Instead of replaying exemplars at a fixed cadence, rehearsal bursts are scheduled dynamically from ADWIN alarms so the network only revisits the buffer when forgetting is detected. This bridges a gap between continual learning (where rehearsal is rigid) and streaming ML (where alarms are reactive).
- Dual signal fusion: The detector will listen to both exemplar validation accuracy and task-specific proxy losses, using multi-metric fusion to reduce false positives. Prior iCaRL implementations rely on a single accuracy signal.
- Energy-aware evaluation: Experiments will not only report accuracy/forgetting but also GPU time and estimated energy cost per task. Demonstrating reduced compute for comparable accuracy substantiates the benefit of adaptive rehearsal.
- Open-source tooling: The codebase will expose modular hooks (detector API, rehearsal scheduler, logging dashboards) so other continual-learning researchers can plug in alternative detectors or buffers.
These additions ensure the work extends beyond reproducing iCaRL or ADWIN individually and documents genuine semester-long exploration.
| Component | Strengths | Limitations |
|---|---|---|
| iCaRL (Incremental Classifier and Representation Learning) | Strong baseline for class-incremental learning, exemplar management, balanced fine-tuning. | Fixed rehearsal schedule increases training time and energy use even during stable phases. |
| ADWIN (Adaptive Windowing) Drift Detector | Provides theoretical guarantees on detecting distribution change in streaming settings, efficient online updates. | Requires performance signals (loss/accuracy) and can trigger false alarms without calibration. |
- Complementary Signals – iCaRL supplies exemplar buffers and evaluation hooks; ADWIN monitors metrics to decide when rehearsal is necessary.
- Resource Awareness – Adaptive triggering reduces redundant rehearsal epochs, aligning with compute-constrained deployment scenarios.
- Clear Research Questions – How sensitive must the detector be? What latency is acceptable between drift detection and recovery? Can dynamic rehearsal match iCaRL accuracy with fewer updates?
| Timeline (2025) | Focus | Outcomes |
|---|---|---|
| Weeks 1-2 (July) | Deep dive into continual learning fundamentals; reproduce simple rehearsal baselines (e.g., ER, GEM). | Literature notes, baseline scripts, evaluation harness (Split CIFAR-10/100). |
| Weeks 3-4 (Aug) | Implement and validate vanilla iCaRL; document architecture, exemplar selection, incremental training loop. | Verified iCaRL baseline with metrics + reproducible notebook. |
| Weeks 4-6 (Sept) | Study ADWIN / drift detection; build lightweight monitor pipeline over rehearsal metrics; run ablation to calibrate thresholds. | Modular ADWIN monitor, experiments on synthetic drift streams. |
| Weeks 7-8 (Sept) | Integrate adaptive trigger into iCaRL loop; design experiments comparing constant vs adaptive rehearsal. | Prototype "Smart Rehearsal" model, initial comparison plots. |
| Weeks 9-10 (Oct) | Optimise, evaluate on additional datasets (e.g., Split Tiny-ImageNet); prepare visualisations and documentation. | Final metrics table, compute usage analysis, polished charts. |
| Week 12 (Nov) | Finalise report, presentation, code clean-up, reproducibility checklist. | Submission-ready artefacts and presentation deck. |
- Data Stream Intake → Tasks arrive sequentially (e.g., Split CIFAR-10).
- Feature Extractor & Classifier (iCaRL backbone) → Train incrementally on current task with exemplar rehearsal buffer.
- Performance Monitor → Maintain validation accuracy / loss on exemplar set.
- ADWIN Drift Detector → Consume performance stream, raise alarms on significant drops.
- Adaptive Rehearsal Trigger
- If no drift: continue lightweight rehearsal (minimal or zero replay).
- If drift detected: launch focused rehearsal burst (balanced fine-tuning + exemplar updates).
- Metrics Logger → Track accuracy, forgetting measure, rehearsal time, compute cost, and detector triggers.
- Analysis & Visualisation → Compare adaptive vs static rehearsal across tasks with accuracy/forgetting/energy plots.
This flow will be implemented modularly so each component can be evaluated independently during development.
| Module | Description | Status (Planned Completion) |
|---|---|---|
| Baseline Core | PyTorch iCaRL backbone with exemplar memory and balanced fine-tuning. | Weeks 3-5 |
| Performance Streamer | Lightweight service to log validation accuracy/loss to ADWIN. | Weeks 6-7 |
| Drift Detector Wrapper | Calibrated ADWIN thresholds, multi-metric fusion, alarm debouncing. | Weeks 7-8 |
| Adaptive Scheduler | Policy translating alarms into rehearsal burst parameters (epochs, buffer refresh). | Weeks 9-10 |
| Experiment Harness | Scripts for Split CIFAR-10/100, Tiny-ImageNet, energy logging, ablations. | Weeks 10-12 |
| Dashboard & Reports | Visual analytics, LaTeX report templates, reproducibility checklist. | Weeks 12-15 |
This table doubles as a progress tracker; each module will produce intermediate artefacts (notebooks, scripts, plots) to show steady semester-long work.
- Problem Statement & Motivation – Written summary of catastrophic forgetting and inefficiencies in static rehearsal.
- Literature Review Snapshot – Two-page synthesis of rehearsal and drift-detection approaches with identified gap.
- System Design Draft – Architecture diagram & flow description (Section 4) plus success criteria.
- Baseline Plan – Experimental setup for iCaRL baseline, datasets, evaluation metrics, and energy-measurement protocol.
- Expected Outcomes – Hypothesised benefits (accuracy parity, reduced rehearsal cost) and risk analysis.
- Implementation Report – Detailed methodology, modular design, and final algorithm.
- Experimental Results – Tables/plots comparing adaptive vs static rehearsal, including compute metrics and alarm statistics.
- Ablation & Sensitivity Analysis – Impact of detector parameters, rehearsal burst size, buffer limits.
- Repository Deliverables – Clean codebase, reproducible scripts, README updates, final presentation slides.
- Reflection & Future Work – Lessons learned, limitations, and potential extensions (e.g., other detectors, task-free settings).
- Rebuffi, S. A., Kolesnikov, A., Sperl, G., & Lampert, C. H. (2017). iCaRL: Incremental Classifier and Representation Learning. CVPR.
- van de Ven, G. M., & Tolias, A. S. (2019). Three scenarios for continual learning. arXiv:1904.07734.
- Bifet, A., & Gavalda, R. (2007). Learning from time-changing data with adaptive windowing. SDM.
- Montiel, J., Read, J., Bifet, A., & Abdessalem, T. (2021). River: Machine learning for streaming data in Python. JMLR.