Energy Efficient Vector Symbolic Architecture Using Spiking Neural Networks
Military / Coalition Issue
Edge of network coalition operations, particularly in the context of IoBT operations, are often performed in energy constrained environments where savings in computation energy efficiency can become the limiting factor in determining where to deploy the sensors and services. Emerging Spiking Neural Network devices are extremely energy efficient but require a new programming paradigm.
Is it possible to use SNN technologies to perform vector process operations and therefore to achieve the desired energy savings?
Core idea and key achievements
Our work using symbolic semantic vectors has been focussed on the use of high dimensional binary vectors to perform the required vector binding and bundling operations. We have shown that it is possible to use SNN devices to perform the required vector processing operations on these vectors but because of the spike density there is little energy saving to be gained. A key achievement is to show that it is possible to perform equivalent operations using sparse vector representations which are more amenable since the equivalent spike density is reduces by a factor of at least x10 on what are already energy efficient devices.
Implications for Defence
The development of SNN devices and their use in energy constrained environments (e.g., drones, edge devices) is a potential disruptive technology that needs to be exploited for defence applications. Our use case for these types of devices was to perform efficient vector processing and our work has illustrated that there are many more potential applications where they can be exploited. Particularly in the areas of AI and machine learning.
Readiness & alternative Defence uses
A number of experimental SNN devices have been developed by organisations such as IBM and Intel and these devices have been shown to operate at significantly lower power than traditional microprocessor architectures (typically x100-x1000 more energy efficiency). The challenge is how to process the devices to perform the required operations. These devices have been successfully demonstrated to be capable of performing energy efficient image processing, and we have shown in our work, they can be used for symbolic vector processing. These types of devices will start to become commercially available in the near term and are ideally suited to the low energy requirements for future IoBT/Iot operations.
Resources and references
- Roy, Deboleena, Priyadarshini Panda, and Kaushik Roy. “Synthesizing images from spatio-temporal representations using spike-based backpropagation.” Frontiers in neuroscience 13 (2019): 621.
- Srinivasan, Gopalakrishnan, and Kaushik Roy. “Restocnet: Residual stochastic binary convolutional spiking neural network for memory-efficient neuromorphic computing.” Frontiers in neuroscience 13 (2019): 189.
- Frontiers in Neuroscience 2020: “Event-driven Backpropagation for Spiking Neural Networks: Enabling Spike-based Learning in State-of-the-art Deep Architectures.”
IBM Europe, Cardiff University, Purdue University