ºÚÁϸ£ÀûÍø

Opportunity

In high-tech equipment, achieving high-throughput and robust capabilities is challenging since they require high-speed perception and low-latency control. State-of-the-art solutions use neural networks for achieving robust and accurate perception and control. However, those solutions hinder the speed and throughput of high-tech equipment. Event-based perception and neuromorphic computing paradigms open up the opportunity to achieve low-latency, high-speed and robust control systems significantly improving the state-of-the-art high-tech equipment. In particular, Spiking Neural Networks (SNNs)  are suitable for targeting neuromorphic sensing and computing systems. The unique feature of SNNs is the use of spiking neurons, which brings the dimension of time into the network’s functioning. The temporal nature of the spikes makes SNNs efficient, as neurons only process and transmit binary information when necessary. Consequently, SNNs can offer low-latency, energy-efficient functional approximation¹. This is in stark contrast to the high computational demands of traditional deep networks, which involve large memory- and compute-intensive matrix multiplications and communication of non-spiking real-valued signals.

In summary, SNNs can bring enormous benefits to the world of real-time control potentially offering extreme-low latency and ultra-high speed for high-performance control systems. This potential of SNNs, as an alternative to traditional deep networks, is due to key properties making them suitable for realizing real-time control systems: 

• Firstly, SNNs are event-driven, which means these systems communicate and compute by means of binary pulses making it efficient in terms of computation and communication.
• Secondly, SNNs are compatible with bio-inspired local learning rules that make the system adaptable to changes over several time-scales including fast changes (μs time scale) enabling high-speed control and adaption. 
• Thirdly, they are asynchronous in nature, which means computation is performed at the speed of the silicon technology instead of the rate of a clock. This makes SNNs extremely low-latency and highly scalable as the signals propagate using handshaking mechanism with delay insensitive designs rather than being valid only during a clock tick. 

These factors alone have led some researchers to deem the exploration of SNNs’ applicability in control tasks as critically important^2,3.

Roadblock

When designing SNN-based closed-loop perception and control systems, we still lack a formal demonstration that shows superior performances in terms of stability, accuracy and speed of SNN-based control solutions. A fundamental theory underpinning SNN-based emulation or direct SNN-based control system design is lacking. Clearly, just as any form of neural networks, SNNs have, in principle the ability to learn the models of the plant (indirect approach) to control, or directly learn the control law (direct approach) using data-driven methods. However, as SNNs process and represent information in a fundamentally different way than traditional neural networks, attempting to use the same learning mechanisms of traditional deep neural networks falls shorts and does not exploit the full capabilities of SNNs and SNN-based control systems⁴.

Research Hypothesis

We aim to translate the properties of spike-based systems into extremely high-rate periodic signals but with low data rate. This can be exploited to achieve stable, high-speed and high-performance, SNN-based perception and closed-loop control systems. To make extreme-low latency and ultra-high speed real-time control of high-tech systems a reality, we need to develop a formal design and learning framework integrating:

• Obj-1) Low-latency SNN-based sensor processing and perception techniques for high-rate periodic signals but with low-data rate.
• Obj-2) SNN-based control algorithm analysis and design theories for high-sampling rate systems but with sparse processing based on binary spikes.
• Obj-3) SNN hardware and software co-design tools for scalable deployment of SNN controllers resulting in compelling ASIC devices.

Solution and research questions

The specific research directions followed to achieve the 3 objectives are the following:

• Res-1) As high-speed control requires robust event-driven perception, only at the time of sensory input and encoded in the time of the event by means of a single spike, alongside high-speed traditional sensors such as encoders⁵, we will develop fusion algorithms that efficiently combine event-based (spiking) data with time-based data to provide excellent information for control.
• Res-2) We develop synthesis tools for spike-based control algorithms exploiting design tools for event-based control^6,7 and hybrid systems analysis⁸.
• Res-3) We co-design hardware and algorithms within a novel framework for scalable deployment of event-trigger controllers in spiking neural networks. We will do so by means of mixed-signal design that results in low-latency and low-power aVLSI systems⁹.

REFERENCES

¹ Gourav Datta and Peter A Beerel. Can deep neural networks be converted to ultra low-latency spiking neural networks? In Design, Automation & Test in Europe Conference & Exhibition, pages 718–723. IEEE, 2022.
² Zhenshan Bing, Claus Meschede, Florian Rohrbein, Kai Huang, and Alois C Knoll. A survey of robotics control based on learning-inspired spiking neural networks. Frontiers in neurorobotics, 12:35, 2018.
³ Dailin Marrero, John Kern, and Claudio Urrea. A novel robotic controller using neural engineering framework-based spiking neural networks. Sensors, 24(2):491, 2024.
⁴ Jason K Eshraghian, Max Ward, Emre O Neftci, Xinxin Wang, Gregor Lenz, Girish Dwivedi, Mohammed Bennamoun, Doo Seok Jeong, and Wei D Lu. Training spiking neural networks using lessons from deep learning. Proceedings of the IEEE, 2023.
⁵ Vibhor Jain, Sajid Mohamed, Dip Goswami, Sander Stuijk: DNN-Based Visual Perception for High-Precision Motion Control. ECC 2024: 2010-2016.
⁶ W.P.M.H. Heemels, K.H. Johansson, and P. Tabuada. An introduction to event-triggered and self-triggered control. In IEEE Conference on Decision and Control (CDC) 2012, Hawaii, USA, pages 3270–3285, December 2012.
⁷ K. Scheres, , R. Postoyan, and W.P.M.H. Heemels. Robustifying event-triggered control to measurement noise. Automatica, 2023.
⁸ W.P.M.H. Heemels, B. De Schutter, J. Lunze, and M. Lazar. Stability analysis and controller synthesis for hybrid dynamical systems. Philosophical Transactions of the Royal Society A, 368:4937–4960, October 2010.
⁹ J Stuijt, M Sifalakis, A Yousefzadeh, F Corradi μBrain: An event-driven and fully synthesizable architecture for spiking neural networks, Frontiers in Neuroscience, 2021.