Replay4NCL: An Efficient Memory Replay-based Methodology for Neuromorphic Continual Learning in Embedded AI Systems

TL;DR

This paper tackles neuromorphic continual learning (NCL) for embedded AI systems, where memory replay methods incur high latency and energy when long timesteps are used. It introduces Replay4NCL, which compresses latent replay data and replays it at small timesteps, with compensatory adjustments to neuron threshold $V_{thr}$ and learning rate $\eta$, plus a latent data insertion strategy. On the SHD class-incremental benchmark, Replay4NCL achieves a Top-1 accuracy of 90.43% (vs 86.22% for the state-of-the-art SpikingLR), while delivering a 4.88× latency reduction, about 20% latent memory savings, and roughly 36% energy savings. These results demonstrate that efficient latent replay enables practical NCL for tightly constrained embedded systems.

Abstract

Neuromorphic Continual Learning (NCL) paradigm leverages Spiking Neural Networks (SNNs) to enable continual learning (CL) capabilities for AI systems to adapt to dynamically changing environments. Currently, the state-of-the-art employ a memory replay-based method to maintain the old knowledge. However, this technique relies on long timesteps and compression-decompression steps, thereby incurring significant latency and energy overheads, which are not suitable for tightly-constrained embedded AI systems (e.g., mobile agents/robotics). To address this, we propose Replay4NCL, a novel efficient memory replay-based methodology for enabling NCL in embedded AI systems. Specifically, Replay4NCL compresses the latent data (old knowledge), then replays them during the NCL training phase with small timesteps, to minimize the processing latency and energy consumption. To compensate the information loss from reduced spikes, we adjust the neuron threshold potential and learning rate settings. Experimental results on the class-incremental scenario with the Spiking Heidelberg Digits (SHD) dataset show that Replay4NCL can preserve old knowledge with Top-1 accuracy of 90.43% compared to 86.22% from the state-of-the-art, while effectively learning new tasks, achieving 4.88x latency speed-up, 20% latent memory saving, and 36.43% energy saving. These results highlight the potential of our Replay4NCL methodology to further advances NCL capabilities for embedded AI systems.

Figures (13)

• Figure 1: (a) Accuracy of the baseline network that faces CF issues in the class-incremental learning scenario. Baseline network refers to an SNN model without any NCL capabilities. (b) Illustration of the use-case of an embedded AI system that requires energy-efficient CL capabilities, e.g., a mobile agent.
• Figure 2: (a) The state-of-the-art memory replay-based work, SpikingLR bib302, incurs significant latency and energy overheads as compared to the baseline network without NCL techniques. (b) An aggressive timestep reduction (e.g., from 100 to 20 timesteps) may lead to significant accuracy degradation.
• Figure 3: An overview of our novel contributions; highlighted in green.
• Figure 4: Overview of our Replay4NCL methodology, showing its key steps such as timestep optimization, parameter adjustments, and data insertion strategy; the novel contributions are highlighted in green.
• Figure 5: Overview of (a) the forward pass, and (b) the backward pass with SG learning technique under supervised settings.