EGGROLL Trains Spiking Neural Networks with 2.23x Speedup
A new gradient-free method slashes memory and time for brain-like AI chips.
Spiking neural networks promise massive energy efficiency on neuromorphic hardware, but training them is notoriously difficult. The discrete spike threshold makes standard backpropagation impossible, while surrogate-gradient methods require external GPUs and cannot run on-chip. Evolution strategies offer a gradient-free alternative, but their memory cost scales with the number of parameters – impractical for large networks.
Researchers from the University of Maryland present EGGROLL, a low-rank factorization of evolution-strategy perturbations that cuts per-generation memory from O(mn) to O(r(m+n)). Tested on a Leaky Integrate-and-Fire SNN with the N-MNIST dataset, EGGROLL achieves 79.21% test accuracy while reducing wall-clock time by 2.23x compared to full-rank ES. The paper demonstrates a clear accuracy–speed tradeoff and positions EGGROLL as a viable path to on-chip learning without surrogate gradients.
- Memory reduced from O(mn) to O(r(m+n)) via low-rank factorization of ES perturbations.
- Achieved 79.21% test accuracy on the N-MNIST dataset.
- 2.23x wall-clock time speedup over full-rank evolution strategies.
Why It Matters
Enables efficient on-chip learning for neuromorphic hardware, advancing energy-efficient AI without reliance on surrogate gradients.