Optimal pruning schedules and thresholds for edge-deployable Tri-Memory continual learning

Determine optimal pruning frequency and aggressiveness for adaptive synaptic pruning in the Tri-Memory Continual Learning architecture deployed on edge devices to maximize adaptability and retention, and develop strategies to dynamically set lower pruning thresholds at the weight, neuron, or module level based on observed usage statistics, device resource constraints, and task complexity.

Background

The paper proposes a neuroscience-inspired Tri-Memory system (Short-Term, Long-Term, and Permanent Memory) with offline adaptive pruning guided by usage statistics to maintain sparsity and resource efficiency on edge devices. While pruning is central to reclaiming capacity and preventing model bloat, the authors identify uncertainty around how frequently and aggressively to apply pruning and how to set thresholds that balance retaining important knowledge with enabling plasticity for new learning.

This question is critical for real-world deployment where compute, memory, and energy are constrained, and improper pruning can cause excessive forgetting or insufficient capacity recovery. The problem asks for principled, adaptive scheduling and thresholding policies that respond to operational conditions and data novelty.