Core Concepts
Efficient Continual Learning in Deep State-Space Models through Regularization-Based Methods.
Abstract
Deep state-space models (DSSMs) are enhanced with continual learning capabilities to adapt to evolving tasks without catastrophic forgetting. Various regularization-based methods are integrated into DSSMs to ensure efficient updates and address memory constraints. Experiments on real-world datasets demonstrate the efficacy of CLDSSMs in overcoming catastrophic forgetting while maintaining computational and memory efficiency.
Stats
The autodifferentiable EnKF has a memory cost of O(dzN) and a computational cost of O(dzdxN).
EWC, MAS, and SI have consistent memory costs of O(dz^2), while LwF incurs a memory cost of O(dzT).
Computational costs for EWC and MAS scale as O(dz^2), SI scales as O(dz^3), and LwF is dominated by the computation of MSE part.
Memory costs for all regularization methods are independent of the number of tasks and size of training data.
Quotes
"Our proposed CLDSSMs integrate mainstream regularization-based continual learning (CL) methods, ensuring efficient updates with constant computational and memory costs for modeling multiple dynamic systems."
"While various competing CL methods exhibit different merits, the proposed CLDSSMs consistently outperform traditional DSSMs in terms of effectively addressing catastrophic forgetting."
"The results highlight the versatility of the CLDSSMs and their applicability to various real-world dynamic system modeling applications."