MPCS: Neuroplastic Continual Learning via Multi-Component Plasticity and Topology-Aware EWC

Problem
This paper addresses the challenge of continual learning systems, which must balance the dual requirements of plasticity (the ability to acquire new knowledge) and stability (the retention of previously learned knowledge). The authors propose a novel architecture, MPCS (Multi-Plasticity Continual System), which integrates multiple mechanisms to enhance neuroplasticity. This work is presented as a preprint and has not yet undergone peer review.

Method
MPCS incorporates eleven complementary mechanisms:

1. Task-driven neurogenesis
2. Fourier-encoded inputs
3. Elastic Weight Consolidation (EWC) regularization
4. Meta-replay
5. Mixed consolidation
6. Hybrid gating
7. Synapse pruning/regeneration
8. Hebbian updates
9. Task similarity routing
10. Adaptive growth control
11. Continuous neuron importance tracking

The architecture is evaluated on MEP-BENCH, a multi-track benchmark comprising 31 tasks across regression, classification, logic, and mixed domains. The evaluation employs a three-dimensional Pareto criterion that assesses task performance (Perf), representation diversity (RD), and gradient conflict rate (GCR). The authors conduct experiments across 15 ablation configurations (3 seeds, 4 tracks, 2000 epochs) to analyze the contributions of each mechanism.

Results
MPCS achieves a Normalized Efficiency Score (NES) of 94.2, positioning it on the Pareto frontier among 9 of 14 gate-passing systems. Key findings include:

• The removal of Fourier encoding results in a 30.7 percentage point drop in performance and a failure rate of 14% on the MEP gate across tasks.
• Global EWC negatively impacts performance (NES = -4.2), while topology-local EWC mitigates this penalty (NES improves from 90.5 to 91.8).
• The highest-performing configuration, MPCS_EFFICIENT, omits EWC entirely, demonstrating a monotonic relationship in high task-similarity scenarios (s_bar ≈ 0.95): global EWC < topology-local EWC < no EWC.
• Removing the Pareto-dominated components (EWC and Hebbian) leads to MPCS_EFFICIENT, which enhances performance by 0.6 percentage points at a significantly reduced compute cost (127 vs. 602 minutes).

Limitations
The authors acknowledge that while topology-local EWC reduces the performance penalty associated with global EWC, it does not eliminate it entirely. Additionally, the study’s reliance on a specific benchmark (MEP-BENCH) may limit the generalizability of the findings. The impact of the various mechanisms in different task domains and their scalability in real-world applications remain unaddressed.

Why it matters
The implications of this work are significant for the field of continual learning, particularly in developing systems that can efficiently balance the trade-off between learning new tasks and retaining old knowledge. The identification of critical components, such as Fourier encoding, and the establishment of a Pareto frontier as a model-compression guide provide actionable insights for future research. This framework could inform the design of more efficient continual learning architectures, potentially leading to advancements in various applications, including robotics, natural language processing, and adaptive systems.

Authors: Joern Hentsch
Source: arXiv:2605.02509
URL: https://arxiv.org/abs/2605.02509v1