Neuromorphic-Based Signal Processing

Modern sensors and detectors provide a high density of channels with resulting large data rates. Offline processing is not feasible due to limited data bandwidth and power consumption for the transport, hence an online classification needs to be done as early as possible. We investigate how neuromorphic methods can assist in here and enable novel edge computing systems based on FPGAs
and custom integrated circuits.
Contact
Neuromorphic Methods and Hardware for Signal Processing on the Edge
High data rate of modern sensors and detectors results in processing and storing large amounts of data with high power consumption and cost. The challenge of processing data of modern sensors can be handled in a strategy of sensor near data processing to refine the data as soon as it is created and reducing the volume by transmitting only information that is needed on a higher level. Event-driven, asynchronous processing as realized in e.g. Spiking Neural Networks (SNNs) as a neuromorphic method to pick out signal patterns from the noise without using synchronized (and therefore energy inefficient) processes. High parallelism is a necessary technique for processing high data rates with low latency, notably present in FPGAs. Chips implementing neuromorphic paradigms are currently developed. Prototyping and accelerating algorithms needed for these data processing on flexible prototypes such as FPGAs aids in this effort.
The SHiP Experiment and the SBT Readout
The Search for Hidden Particles (SHiP) is a fixed target experiment proposed at the CERN Super Proton Synchrotron (SPS) accelerator to search for long-lived exotic particles associated with Hidden Sectors and Dark Matter [1].
The decay volume between the detectors is covered by the Surrounding Background Tagger (SBT), which is based on liquid scintillator (LS) cells. The purpose of the SBT is to efficiently detect particles leaving or entering the decay vessel outside of SHiP detection solid angle [2,4].
The SBT is partitioned into ~900 LS cells of slightly varying geometry, which are read out by two sets of 40 silicon photomultipliers (SiPM) in order to increase detection efficiency and reconstruct the position of the particle impact in the cell [4].



The SiPMs are amplified by a dedicated chip resulting in eight signals from the SiPMs groups as well as two analog sum signals of two complementary gains of all 40 SiPMs These signals represent both high and low light yield events.
The two analog sums have to be digitized at sufficiently high sensitivity and sampling rates of up to 800 MSPS within the readout electronics for the SHiP’s SBT to be self-triggering and charactering the particle impact events sufficiently.
The data acquisition (DAQ) needs to achieve highest efficiency, being confronted with high mean event rates per cell currently estimated to be in the order of 1 MHz.
Novel Data Reduction Techniques
Classical readout schemes perform thresholding techniques to classify data. This is simple to realize in hardware but lacks performance when signal and noise are hard to distinguish. Neuromorphic signal processing is especially suited for feature extraction in low signal to noise ratios and sparce information processing in noisy invironment. Also, asynchronous data processing is especially suited for time extraction. Hence, novel approaches including neuromorphic computing promise a significant performance improvement in power consumption and time resolution. With this in mind, we plan to implement novel edge computing systems into the readout chain of large scale detectors.
Applications lie in the fields of high-energy physics experiments and medical imaging. Hence, this topic is tightly coupled with our research areas for AI for Imaging and Bio-Inspired Networks & Systems, and Neuromorphic Edge Computing Systems.