spike-based signal processing

Investigation into neuroscience inspired spike-based signal processing has potential to provide new methods for signal processing as well as greater insight into the current learned algorithms used for spike train data. As a secondary benefit it may also improve our current understanding of the neural circuitry that inspires spike-based signal processing. The idea takes its inspiration from the observed behaviour of neural circuits in the brain which code information using the temporal spacing of pulses, each pulse being of the same magnitude. Information processing by the brain is impressive for several reasons:
- Efficiency: The brain uses a fraction of the energy required by computers for comparable tasks.
- Robustness: The constant loss and replacement of neurons in the brain does not affect its ability to process information.
- Performance: The brain still outperforms the state of the art in many common tasks such as speech and face recognition.
These qualities are likely to be due to a multitude of factors at different scales both systemic and physiological, however, it is reasonable to assume that there are advantages to be gained from processing signals in the form of spike trains.
This project will explore and develop the mathematical foundations of spike-based signal processing aligning with the EPSRC research area of digital signal processing and strategic theme of Information and Communication Technology. Most of the research into spike train signals has focused on the analysis of the observed response of neural circuits to stimuli and has not been viewed from an engineering and signal processing perspective. Currently we lack the mathematical fundamentals to perform common operations used in other areas of signal processing on spiking signals. For discrete and continuous time signals, these operations have allowed huge advances in processing signals and allowed us to reason about modern image processing and machine learning. Methods for performing these operations on spike train data would therefore give us greater insight into the advantages, disadvantages and opportunities afforded by processing signals in the form of spike trains. This in turn may be of interest in areas outside of engineering such as neuroscience where a greater intuition about spike train signals would help explain the functionality of neural circuits.
We intend to investigate how to implement the wavelet transform in this new context given the relevance to signal processing of this transform. We are also interested in understanding how the notion of sparsity which is used in many signal processing applications translates to the case of spiking signals. We are also interested in addressing fundamental sampling questions. Specifically, under which conditions it is possible to reconstruct a waveform from the train of spikes.
Spike based signal processing already has practical engineering applications. For example, Dynamic vision sensors in which information about the stimulus is temporally encoded in the output provide much finer time resolution whilst reducing power requirements. This has many possible applications such as in CCTV, drone automation or driverless cars. Driverless vehicles will require ultra-low latency from stimulus to reaction to operate safely. Current high-speed camera technology also has high power requirements that hamper their use on autonomous vehicles where energy consumption is highly constrained. Spike based sensors and signal processing provide a solution to these problems. However, a more rigorous mathematical understanding of this technology is required, particularly given the safety requirements of automated vehicles and other machines.