Mitigating Catastrophic Forgetting in Molecular Property Prediction via Refresh Learning and Pareto Optimization

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Chemistry

Background:

• Continual Learning (CL) is crucial for Large Language Models (LLMs) to adapt to new data without forgetting, especially in dynamic fields like Molecular Property Prediction (MP).
• LLMs often suffer from Catastrophic Forgetting (CF), where new learning erodes prior knowledge, particularly with shifting data distributions in chemical, genomic, and proteomic datasets.
• Existing replay-based CL methods use memory buffers but neglect inter-episode relationships, leading to suboptimal performance on previously encountered data.

Purpose of the Study:

• To develop a novel Continual Learning framework that mitigates Catastrophic Forgetting in Large Language Models for Molecular Property Prediction.
• To propose a unified hierarchical gradient aggregation framework integrating Multi-task Learning (MTL) and Refresh Learning (RL) principles.
• To enhance the stability-plasticity trade-off in sequential learning for LLMs.

Main Methods:

• A new Multi-task Learning framework, MTL-PORL (Multi-task Learner-Pareto Optimized Refresh Learning), was developed using the ChemBERTa model.
• The framework incorporates Refresh Learning (RL), inspired by neuroscience, which discards outdated information to improve retention and new learning.
• Pareto Optimization (PO) and a hyper-gradient approach are employed to unlearn and relearn current data, enhancing existing CL methods as a plug-in.

Main Results:

• MTL-PORL achieved high Anytime Average Accuracy (91.63%-94.89%) and Test Accuracy (92.48%-96.86%) across BBBP, bitter, and sweet datasets.
• The model demonstrated a low Forgetting Measure (-0.0048 to -0.0063), indicating effective knowledge retention.
• Empirical analysis showed significant improvements in sequential learning compared to existing methods, effectively addressing the stability-plasticity dilemma.

Conclusions:

• MTL-PORL offers a significant advancement in Continual Learning for LLMs, particularly for Molecular Property Prediction tasks.
• The proposed framework effectively combats Catastrophic Forgetting and improves performance on dynamic datasets.
• The Refresh Learning approach, combined with Pareto Optimization, provides a flexible and effective solution for enhancing sequential learning in LLMs.