Pseudo-rehearsal: A simple solution to catastrophic forgetting for NLP

2017-08-23 · Source: Explosion · Developer tools and consulting for AI, Machine Learning and NLP - Explosion.ai · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Intermediate, quick

Summary

Pseudo-rehearsal is presented as a straightforward technique to mitigate catastrophic forgetting when fine-tuning pre-trained Natural Language Processing (NLP) models. This problem typically arises when a model is updated for new labels or error correction, causing it to lose previously learned knowledge. The pseudo-rehearsal method addresses this by generating synthetic training examples: the original, pre-trained model is used to label a set of unlabeled data, creating "pseudo-examples." These pseudo-labeled examples are then combined with the new, specific fine-tuning data during the update process. This mixing helps the model retain its original capabilities while adapting to new information, effectively preventing the degradation of performance on previously mastered tasks.

Key takeaway

For NLP Engineers fine-tuning pre-trained models to add new labels or correct errors, you should implement pseudo-rehearsal to prevent catastrophic forgetting. By generating and mixing pseudo-labeled examples from your original model into the fine-tuning updates, you can ensure the model retains its broad capabilities while effectively learning new, specific tasks. This approach minimizes performance degradation on existing knowledge, making your fine-tuning process more robust and efficient.

Key insights

Pseudo-rehearsal prevents catastrophic forgetting by mixing original model-labeled examples during fine-tuning.

Principles

• Catastrophic forgetting occurs during fine-tuning.
• Mixing old and new data preserves knowledge.

Method

Use the original pre-trained model to label examples, then integrate these pseudo-labeled examples with new fine-tuning data during model updates.

In practice

• Apply pseudo-rehearsal when adding new labels.
• Use for correcting specific model errors.

Topics

• Pseudo-rehearsal
• Catastrophic Forgetting
• NLP Fine-tuning
• Pre-trained Models
• Machine Learning

Best for: Machine Learning Engineer, NLP Engineer, AI Scientist

Related on AIssential

• MIT's new fine-tuning method lets LLMs learn new skills without losing old ones — SDFT enables LLMs to learn new skills continually without forgetting old ones, using self-distillation and in-context learning.
• Not Just After One: Sleep-Inspired Replay Prevents Catastrophic Forgetting After Sequential Tasks — Sleep-inspired replay after sequential training prevents catastrophic forgetting in neural networks, enabling continuous learning.
• Forgetting is Not Erasure: Recovering Latent Knowledge via Transport Keys — Catastrophic forgetting often reflects interface drift, not erasure, allowing latent knowledge recovery via alignment.
• Fine-Tuning Regimes Define Distinct Continual Learning Problems — Fine-tuning regimes significantly alter method performance and forgetting in continual learning.
• AI Shadow Brain: More Skills (no .md) — GDIO prevents catastrophic forgetting by expanding MLP dimensions and creating orthogonal knowledge subspaces within LLMs.
• Why Fine-Tuning Encourages Hallucinations and How to Fix It — SFT-induced hallucinations stem from factual forgetting due to semantic interference, mitigable by self-distillation or parameter freezing.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Explosion · Developer tools and consulting for AI, Machine Learning and NLP - Explosion.ai.