Spiking Neural Networks on Unconventional Nanodevices for the Detection of Acoustic Incidents

• Applicant: Professor Dr.-Ing. Stefan E. Schulz
• Subject Area: Electronic Semiconductors, Components and Circuits, Integrated Systems, Sensor Technology, Theoretical Electrical Engineering
• Term: since 2024
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 545923700

Based on a hardware-implemented self-learning artificial neural network, an energy-saving system for event detection in acoustic signals is to be developed. In order to avoid the power-intensive analog-to-digital conversion, large parts of the signal processing are relocated to the analog area. Inspired by the human cochlea, a dynamic analog filterbank extracts spectral features accentuating frequency bands which carry information while pruning irrelevant filters for a given application. The analog filter outputs are converted to spike rates, feeding a fully analog Spiking Neural Network (SNN) with four to six layers, each having up to 128 neurons. Crossbar arrays (CBAs) of interface-type BFO-memristors emulate layers of fully connected synaptic connections between neural layers. By using appropriately shaped spike pulses, local learning algorithms such as spike-timing-dependent plasticity (STDP) will be implemented using the memristors, thereby considering the volatile characteristic of interface type memristors as synaptic forgetting rate. The investigations will focus on the effects of different design parameters, such as the spike waveform and titanium doping of the memristive layer, on the memristive learning function. The SNN is constructed by heterogeneous integration of the synaptic crossbar arrays with CMOS based ultra-low power implementations of leaky integrate-and-fire (LIF) neurons. Both chips are combined through chip-to-board cointegration in the first iteration and through chip-to-wafer heterointegration in the second iteration. The first approach enables flexible connection schemes for prototyping. The second approach aims to minimize the total area of the system by stacking CBA and CMOS on top of each other. Real-time processing in SNNs requires precise timing in the neural circuits. To achieve long time constants, the investigation will explore various methods, including tunable operational transconductance amplifiers, digital counters, and volatile memristors.The targeted application field of the project are self-learning, biologically inspired sensor systems for mobile applications. The project’s demonstrator will be utilized and evaluated for keyword spotting, with a focus on the system’s power usage. In addition, first investigations with photosensitive BFO-CBAs will shed light on potential applications of the developed system with visual input stimuli.

• DFG Programme: Research Grants
• International Connection: France
• Partner Organisation: Agence Nationale de la Recherche / The French National Research Agency
• Cooperation Partner: Professor Wenceslas Rahajandraibe