DFG-ANR: Spiking Neural Networks on Unconventional Nanodevices for the Detection of Acoustic Incidents (Sound-AI)

In this project, we aim to develop a self-learning system for ultra-low-power detection of events in acoustic signals. Usually, an audio signal is regularly sampled and processed digitally. The proposed system shifts large parts of the signal processing to the analog domain in order to avoid power-hungry analog-to-digital conversion which is critical in always-on edge computing applications. Spectral features are extracted by an analog filter-bank implementing pre-accentuation and pre-attenuation dynamic filtering mechanisms as in human cochlea and having the ability of switching on/off any filter in the bench for pruning the irrelevant filters for a given application. The switching filter ability allows to implement an important number of highly selective band-pass filters in the bank (64 or 128), as most of them will be off during operation, giving an extended versatility to the system. Analog parameters tuning during learning and/or dynamically during operation will improve the reliability of the classification by, for example, attenuating frequency bands where an important noise is present, as done in human cochlea. The instantaneous output powers of filters are converted to spike rates which are fed to a fully analog Spiking Neural Network (SNN). Considering the number of band-pass filters in the cochlea, the SNN is composed of 4 to 6 neural layers, each one having at most 128 neurons fully connected to the preceding and next layer. Definition of neural interconnections will be done during the project, but recursive schemes will be explored as well as lateral inhibition. In the SNN, synaptic interconnections are implemented with memristive crossbar arrays (CBAs). Interface-type BFO-memristors have been shown to efficiently emulate Hebbian learning algorithms such as spike timing-dependent plasticity (STDP), which enables the network to become susceptive to repeating temporal patterns in the frequency spectrum of the input signal. For the first time, we suggest exploiting the volatile characteristic of interface-type memristors as the forgetting rate of a synaptic connection. The memristive CBAs will be placed on top of a CMOS-backplane, containing analog neuron circuits which are compatible with the proposed synapses. An embedded processor will interface the SNN and process its output spikes.