Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Continual Learning of Multi-modal Dynamics with External Memory

Abdullah Akgül, Gozde Unal, Melih Kandemir

TL;DR

This work addresses continual learning for dynamical environments where new, unknown modes emerge sequentially and ground-truth mode labels are unavailable. It introduces the Continual Dynamic Dirichlet Process (CDDP), an approach that stores fixed-size mode descriptors in an external neural episodic memory, uses a Dirichlet Process prior to encourage a compact set of modes, and feeds retrieved descriptors into the state-transition dynamics for cross-task transfer. The method blends Bayesian state-space modeling with variational inference and memory-augmented neural networks, providing a principled, unsupervised mechanism for encoding and reusing mode information across tasks. Empirical results on five time-series datasets show that CDDP consistently outperforms a strong VCL-based baseline, highlighting the importance of memory-driven, mode-aware transfer in continual learning for multi-modal dynamics.

Abstract

We study the problem of fitting a model to a dynamical environment when new modes of behavior emerge sequentially. The learning model is aware when a new mode appears, but it cannot access the true modes of individual training sequences. The state-of-the-art continual learning approaches cannot handle this setup, because parameter transfer suffers from catastrophic interference and episodic memory design requires the knowledge of the ground-truth modes of sequences. We devise a novel continual learning method that overcomes both limitations by maintaining a \textit{descriptor} of the mode of an encountered sequence in a neural episodic memory. We employ a Dirichlet Process prior on the attention weights of the memory to foster efficient storage of the mode descriptors. Our method performs continual learning by transferring knowledge across tasks by retrieving the descriptors of similar modes of past tasks to the mode of a current sequence and feeding this descriptor into its transition kernel as control input. We observe the continual learning performance of our method to compare favorably to the mainstream parameter transfer approach.

Continual Learning of Multi-modal Dynamics with External Memory

TL;DR

This work addresses continual learning for dynamical environments where new, unknown modes emerge sequentially and ground-truth mode labels are unavailable. It introduces the Continual Dynamic Dirichlet Process (CDDP), an approach that stores fixed-size mode descriptors in an external neural episodic memory, uses a Dirichlet Process prior to encourage a compact set of modes, and feeds retrieved descriptors into the state-transition dynamics for cross-task transfer. The method blends Bayesian state-space modeling with variational inference and memory-augmented neural networks, providing a principled, unsupervised mechanism for encoding and reusing mode information across tasks. Empirical results on five time-series datasets show that CDDP consistently outperforms a strong VCL-based baseline, highlighting the importance of memory-driven, mode-aware transfer in continual learning for multi-modal dynamics.

Abstract

We study the problem of fitting a model to a dynamical environment when new modes of behavior emerge sequentially. The learning model is aware when a new mode appears, but it cannot access the true modes of individual training sequences. The state-of-the-art continual learning approaches cannot handle this setup, because parameter transfer suffers from catastrophic interference and episodic memory design requires the knowledge of the ground-truth modes of sequences. We devise a novel continual learning method that overcomes both limitations by maintaining a \textit{descriptor} of the mode of an encountered sequence in a neural episodic memory. We employ a Dirichlet Process prior on the attention weights of the memory to foster efficient storage of the mode descriptors. Our method performs continual learning by transferring knowledge across tasks by retrieving the descriptors of similar modes of past tasks to the mode of a current sequence and feeding this descriptor into its transition kernel as control input. We observe the continual learning performance of our method to compare favorably to the mainstream parameter transfer approach.
Paper Structure (29 sections, 14 equations, 2 figures, 2 tables)

This paper contains 29 sections, 14 equations, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Continual Multi-Modal Dynamics Learning
  3. Novel Baseline: Variational Continual Learning for Bayesian State-Space Models
  4. Bayesian State Space Models (BSSM):
  5. Variational Inference
  6. VCL for BSSMs
  7. Target Model: The Continual Dynamic Dirichlet Process
  8. Dirichlet Processes
  9. Neural Episodic Memory.
  10. The full model.
  11. Inference.
  12. Prediction.
  13. Related Work
  14. State Space Models.
  15. Attention and Memory in Neural Nets.
  16. ...and 14 more sections

Figures (2)

  • Figure 1: Our research shows that inferring and using mode descriptors from observed sequences significantly enhances CL performance. Storing these descriptors in external neural episodic memory and integrating them into subsequent tasks drives improved learning outcomes.
  • Figure 2: After learning is finished, CDDP and VCL-BSSM predictions on Character Trajectories are displayed alongside ground-truth trajectories. Context and predictions are plotted in red and blue respectively. CDDP exhibits superior memory of the initial task and precise adaptation to the last task due to its external episodic memory.