"Gradient Localization Improves Lifelong Pretraining of Language Models"

Smart layer targeting helps LLMs update their knowledge like humans do

Traced Gradient Layers (TGL) lets LLMs learn new facts without forgetting old ones by targeting specific neural layers.

LLMs store knowledge in their parameters but struggle to update this knowledge over time without catastrophic forgetting. This paper discovers that different types of knowledge are stored in specific layers, and proposes a method to target these layers during updates.

-----

https://arxiv.org/abs/2411.04448

🤔 **Original Problem**:

LLMs need constant knowledge updates but current methods either fail to learn new information or forget previously learned knowledge. The core challenge is understanding how different types of knowledge are stored within the model's parameters.

-----

🔧 **Solution in this Paper**:

→ The paper introduces Traced Gradient Layers (TGL), which identifies specific layers responsible for storing different types of knowledge.

→ TGL analyzes gradient patterns when processing temporal knowledge versus regular pretraining data.

→ The method either freezes parameters or applies adaptive learning rates to specific layers based on their gradient patterns.

→ This targeted approach helps preserve old knowledge while effectively learning new information.

-----

💡 **Key Insights**:

→ Different types of knowledge are localized to specific sets of parameters within LLMs

→ Sequences with updated entities show larger gradient norms in certain layers

→ Early and middle layers show particularly high activity for temporal knowledge

-----

📊 **Results**:

→ TGL improved performance across all continual learning baselines

→ Reduced perplexity on both new entity recognition (ECBD-NP) and entity relation updates (TempLAMA)

→ Avoided catastrophic forgetting while maintaining model performance