Escaping Stability-Plasticity Dilemma in Online Continual Learning for Motion Forecasting via Synergetic Memory...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Deep neural networks (DNNs) excel in intelligent systems but face catastrophic forgetting.
• This forgetting creates a stability-plasticity dilemma, hindering continuous learning.
• Catastrophic forgetting impairs DNN performance on previously learned tasks after new data adaptation.

Purpose of the Study:

• Propose a novel continual learning (CL) method, Synergetic Memory Rehearsal (SyReM).
• Address the stability-plasticity dilemma in deep neural networks.
• Mitigate catastrophic forgetting while improving learning plasticity.

Main Methods:

• SyReM utilizes a compact memory buffer to store learned knowledge.
• An inequality constraint limits average loss increments over the memory buffer for stability.
• A selective memory rehearsal mechanism enhances plasticity by rehearsing similar data samples based on cosine similarity of loss gradients.

Main Results:

• SyReM significantly mitigates catastrophic forgetting in past scenarios.
• The method improves forecasting accuracy in new scenarios.
• Experiments on naturalistic driving datasets validate SyReM's effectiveness compared to baselines.

Conclusions:

• SyReM effectively balances memory stability and learning plasticity in continual learning.
• The proposed method offers a promising solution for deep neural networks in dynamic environments.
• SyReM demonstrates superior performance in motion forecasting tasks under online CL settings.