Neuromorphic (NM) engineering has been a research topic in academia since the '80s, but only in the last decade this niche field has widened to a much larger research community. From one side, a new renaissance in academic and commercial research in Neural Networks (NN) has shed a new light on the potentials of Spiking Neural Networks (SNN), seen by many as the third generation of NN. From the other side, chip manufacturers have started to look at alternative computing and sensing architecture, due to well-known manufacturing limitations highlighted by Moore's law. Both factors have indeed given rise to the view in the research community and in industry that NM and SNN can provide additional benefits to what is currently available through Deep Learning (DL), in terms of lower SWaP (Size, Weight and Power) profile.

However, although current state of the art in NM and SNN seems to clearly point in that direction, there is still a practical issue with the widespread use of SNN in real scenarios, as a clear killer application for the use of SNN has still not emerged. As a term of comparison, for example, applications like object detection and recognition have determined the massive growth in research and deployment of Deep Neural Networks (DNN) using backpropagation.

Even though academia and industry are fully aware of the potentials and benefits intrinsic to the use of SNN, there are still fundamental questions left unanswered, that in many cases lead to skepticism from users in the deployment of SNN in real applications and scenarios, to the advantage of more widely tested and trusted backpropagation-based DNN. This is certainly the case in industry, where nowadays it is relatively simple for engineers without strong background in DL to find suitable DNN that can be easily put to use with minor tweaking.

This is not the case for SNN. In fact, the first stumbling block in the use of SNN is the relative lack of available NM data, when compared to conventional sensing data used in DL. But more fundamental, there is essentially a debate on best methods and practice to effectively train SNN. Again, for example, in DNN, backpropagation has been investigated in depth and it has proven robust to the point where DNN have grown so much that hardware has struggled to catch up.

Methods that emulates backpropagation in the spiking domain are available, but some researchers argue that one of the fundamental advantages of SNN is in their ability to extract temporal information from the data in an unsupervised fashion. However, this leads to obvious questions from users on robustness of training approaches and their real feasibility, when it comes to real applications and implementations.

As a follow-on from the previous very successful call for papers on the theoretical advances and practical applications of SNN, this second collection of articles seeks research papers that aim to address the aforementioned issues related to practical application of SNN, focusing on:

• effective methodologies for unsupervised training of SNN on larger and more complex networks

• continual learning in SNN trained with backpropagation-based methods

• practical implementation of SNN in real life applications and scenarios

• fundamental applications that fully exploit all benefits of SNN

• quantitative evaluations of SWaP profile reductions in SNN-based applications and implementations

• effective and efficient implementation of SNN on conventional (non-NM) computing architectures

• SNN-based processing leveraging event-based sensing

This Research Topic is part of the Theoretical Advances and Practical Applications of Spiking Neural Networks, Volume I and Volume II series. Link to Volume I:

Theoretical Advances and Practical Applications of Spiking Neural Networks, Volume I

Submit your idea

Impact

Frontiers in Neuroscience

Submit your idea

Impact