CLP-SNN: Spiking Neural Network for Online Continual Learning on Intel Loihi 2
A study introduces CLP-SNN, a spiking neural network developed in collaboration with Intel's Loihi 2 neuromorphic processor, aimed at facilitating online continual learning on edge devices. This network employs a self-normalizing local learning mechanism and a spike-driven neural state machine, allowing for autonomous on-chip learning while preventing catastrophic forgetting. In few-shot experiments on OpenLORIS, CLP-SNN achieves replay-based accuracy without the need for rehearsal. When tested on Loihi 2, it demonstrates 113 times lower latency (0.33 ms compared to 37.3 ms) and 6,600 times lower energy consumption (0.05 mJ versus 333 mJ) than the leading edge-GPU baseline. These improvements stem from algorithmic efficiency and neuromorphic hardware co-design, showcasing the promise of brain-inspired algorithms in energy-efficient AI systems.
Key facts
- CLP-SNN is a spiking neural network for online continual learning.
- It is implemented on Intel's Loihi 2 neuromorphic processor.
- CLP-SNN uses a self-normalizing local learning rule and a spike-driven neural state machine.
- On OpenLORIS few-shot experiments, CLP-SNN matches replay-based accuracy without rehearsal.
- On Loihi 2, CLP-SNN achieves 113x lower latency (0.33 ms vs. 37.3 ms) than edge-GPU.
- On Loihi 2, CLP-SNN achieves 6,600x lower energy (0.05 mJ vs. 333 mJ) than edge-GPU.
- Algorithmic efficiency contributes ~14.5x latency and ~22.6x energy improvement on the same GPU.
- Neuromorphic hardware co-design contributes ~7.8x latency and ~295x energy improvement.
Entities
Institutions
- Intel