概述
Continual-learning methods attempt to balance adaptation and retention.
深入探讨
Catastrophic forgetting occurs when learning new tasks degrades earlier capabilities. EWC penalizes changes to parameters important to earlier tasks; GEM constrains gradients using episodic examples under a local first-order approximation; iCaRL stores exemplars and uses class-mean features. LoRA freezes base weights and learns low-rank task updates, which may reduce direct interference but does not guarantee retention. The original RAG paper distinguishes sequence-level and token-level document marginalization. These mechanisms address different issues and require evaluation on old and new tasks.
现实世界的实施
Elastic Weight Consolidation penalizes changes to parameters estimated to be important for earlier tasks.
Experience replay mixes stored earlier examples with new-task examples during training.
LoRA keeps task-specific low-rank updates separate from frozen base weights in some adaptation settings.
Retrieval-augmented generation can consult an external index rather than storing every fact only in model parameters.
灾难性遗忘的未来
Evaluate old and new task performance after each update and make the retention criterion explicit. Choose methods based on data availability, storage and privacy constraints. Research benchmarks do not establish that any single technique prevents forgetting in all deployment settings.
付诸实践
为什么这很重要
- 速度与规模. 语言工作流程可以在不牺牲一致性的情况下更快地移动。
- 交通与覆盖范围. 它扩展了跨语言和沟通方式的访问。
- 更清晰的判决. 团队可以花更多时间进行判断，而自动化则可以处理重复。
留意
- 幻觉的事实可以悄悄地进入报告、支持流程或研究成果。
- 及时的敏感性可能会在类似的请求中产生不一致的结果。
- 如果访问控制薄弱，敏感文本数据可能会暴露。
后续步骤
- 在推出之前定义输出格式、语气和质量标准。
- 当准确性很重要时，请使用可信来源进行地面响应。
- 为高风险输出保留人工审查检查点。
- 跟踪故障模式并定期重新训练提示或工作流程。
不断探索
常见问题
什么是灾难性遗忘？
Catastrophic forgetting occurs when training on new tasks degrades performance on previously learned tasks. Continual-learning methods attempt to balance adaptation and retention.
哪种情况描述了神经网络中的灾难性遗忘？
灾难性遗忘是指当网络接受新任务训练时，旧功能突然丧失，因为共享权重被覆盖。
为什么神经网络会发生灾难性遗忘？
知识存在于共享参数中，因此对新数据的训练会改变相同的权重并消除早期的学习。
弹性权重巩固（EWC）如何减少遗忘？
EWC 添加了正则化惩罚，减慢了对旧任务重要的参数的更新速度，这些参数是通过 Fisher 信息估计的。
体验重播（排练）如何对抗遗忘？
排练混合过去的示例（存储或生成），同时学习新任务，从而加强旧的表现。
What does LoRA do to base-model weights when adding a task-specific adapter?
LoRA freezes base weights and trains low-rank updates for a task. This can reduce direct parameter interference but does not guarantee retention across every continual-learning sequence.