AI accelerationists frequently hype the amazing increases in productivity that AI will deliver, and the enormous benefits to society that will result.

It’s such a privilege to set foot on this ancient land. These rocks have seen tens of thousands of years of First Nations communities — “Australia” is something brand new, in comparison, and we have the moral obligation to listen and learn, and correct our mistakes.

As a result, we use *much* less implementation resources for beamforming than standard approaches. Using an SNN means we are also very power efficient, while still achieving state-of-the-art accuracy for SSNs, comparable with standard super-resolution methods such as MUSIC [3]!

By using the Hilbert Transform we developed a single beamforming approach that works well in the narrowband case, and can use all frequencies of a wideband signal to work well in the wideband case!

We took a different approach, designed for arrays with many microphones (>2). We start with a construct called the Hilbert Transform to estimate the phase of each signals and to encode the phases as spikes. We then use a beamforming method to estimate the source direction.

Most SNN implementations of sound source localization take this approach, using the precise differences in spike times generated by a single-frequency sound at two microphones to estimate the location of an audio source.

Mammals use the fact that audio sources from different directions lead to very precise differences in arrival time between our two ears — known as inter-aural time differences (ITDs). ITDs are encoded by the differences in neuronal spike times produced by our cochleas. [2]

We've built a new system for sound source localization, based on spiking neural networks (SNNs), that sets a new state-of-the-art for SNN implementations, is extremely power efficient, and even matches the accuracy of standard DSP-based approaches! [1] arxiv.org/abs/2402.11748

Sound source localization is an important part of dealing with audio. People use it to help pay attention to someone talking to us in a noisy environment. Smart home speakers use it to identify when someone is speaking, to focus on their voice and reject the background noise.

#Benchmarking is crucial for #Neuromorphic #processors to reach mainstream commercial acceptance. Most NM benchmarks are one-off, and don't compare results against commodity hardware. The Neurobench project is trying to fix that!

Christmas came early! Probably the densest collection of ultra-low-power Neuromorphic sensory processors on this side of the country 😂

From left to right: XyloAudio 3 (brand new!); stand-alone Speck modules; a Speck HID module (unreleased!); Speck dev kit; XyloIMU dev kit and XyloAudio 2.

Hi! 👋 I am working to make #SpikingNeuralNetworks the next big thing for #MachineLearning. Currently I'm at SynSense, focussed on applications for #SNNs, as well as toolchains to enable ML and SW Engineers to use our #Neuromorphic technology for low-power sensing and processing.