Research Question
How does protection against catastrophic forgetting change as a model learns more tasks?
Why It Matters
A continuously learning system needs to retain earlier knowledge while adapting to new tasks.
My Role
The project work includes a training and evaluation pipeline, comparative experiments, and result visualization.
Methodology
Compare SGD, L2 regularization, and Elastic Weight Consolidation on a sequence of 20 Permuted MNIST tasks.
Results and Evidence
The project records include per-task result tables and forgetting curves. Figures and reproducibility notes will be added here.
Limitations
A benchmark comparison is not evidence of general-purpose lifelong learning. Variation across random seeds also matters.
Reflection and Next Steps
Document the experimental setup and evaluate the stability of the comparison across repeated runs.
研究问题
随着模型学习的任务增多，抵抗灾难性遗忘的效果如何变化？
为什么重要
持续学习系统需要在适应新任务的同时保留旧知识。
个人贡献
项目工作包括训练评估流程、对照实验以及结果可视化。
研究方法
在 20 个连续的 Permuted MNIST 任务上，对比 SGD、L2 正则化与弹性权重巩固（EWC）。
结果与证据
项目记录包含各任务的结果表和遗忘曲线。后续会补充图表与复现说明。
局限
基准实验不等于通用终身学习能力；不同随机种子下的波动同样重要。
反思与下一步
整理实验配置，并通过重复实验检查对照结果的稳定性。