Related Experiment Video
Updated: Jul 7, 2026
07:34
A Simple Stimulatory Device for Evoking Point-like Tactile Stimuli: A Searchlight for LFP to Spike Transitions
Published on: March 25, 2014
9.9K
An exact mapping from ReLU networks to spiking neural networks
Ana Stanojevic1, Stanisław Woźniak2, Guillaume Bellec3
1IBM Research Europe - Zurich, Rüschlikon, Switzerland; École polytechnique fédérale de Lausanne, School of Life Sciences and School of Computer and Communication Sciences, Lausanne EPFL, Switzerland.
Summary
We developed an exact method to convert deep Rectified Linear Unit (ReLU) networks into energy-efficient spiking neural networks (SNNs). This approach achieves zero performance loss, enabling low-power AI applications.
Area of Science:
- Artificial Intelligence
- Computational Neuroscience
- Machine Learning
Background:
- Deep spiking neural networks (SNNs) are promising for low-power AI.
- Training deep SNNs or converting existing networks without performance loss remains a significant challenge.
Purpose of the Study:
- To propose an exact mapping method for converting deep Rectified Linear Unit (ReLU) networks into single-spike SNNs.
- To demonstrate that this conversion can be achieved without any performance degradation.
Main Methods:
- Developed a constructive proof for an exact mapping from multi-layer ReLU networks to SNNs.
- Assumed access to trained ReLU network parameters (weights, biases) and representative training data.
- Applied the mapping to networks with convolutional, batch normalization, and max pooling layers.
Main Results:
- Achieved zero percent accuracy drop on benchmark datasets: CIFAR10, CIFAR100, Places365, and PASS.
- Demonstrated the feasibility of replacing deep ReLU networks with energy-efficient single-spike SNNs.
Conclusions:
- An arbitrary deep ReLU network can be precisely converted into an energy-efficient single-spike SNN.
- This method enables the deployment of high-performance, low-power AI systems without compromising accuracy.
