The last decade has seen a massive renewed interest, both in academia and industry, in Artificial Intelligence and more generally in systems that can mimic humans and learn from experience. This interest has spread also to other fields of research that share points of contact with AI. In this context, Neuromorphic (NM) engineering has definitely benefited from this. Academia and industry are now aware of the existence of Neuromorphic, however there is still an underlying misunderstanding of what NM really is and what it is capable of.NM has been a research topic in academia since the '80s, but only in the last decade this niche research field has widened to a much larger research community. From one side, this interest in NM has shed a new light on the potentials of Spiking Neural Networks (SNN), seen by many as the third generation of Neural Networks. From the other side, chip manufacturers have started to look at alternative computing 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 the 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, i.e. a clear 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 scepticism 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 a 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 fundamentally, there is a debate on best methods and practices 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 has 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.Therefore, in this Research Topic we welcome articles that aim to address the aforementioned issues related to practical application of SNN, specifically focusing on:• effective methodologies for unsupervised training of SNN on larger and more complex networks• 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
The last decade has seen a massive renewed interest, both in academia and industry, in Artificial Intelligence and more generally in systems that can mimic humans and learn from experience. This interest has spread also to other fields of research that share points of contact with AI. In this context, Neuromorphic (NM) engineering has definitely benefited from this. Academia and industry are now aware of the existence of Neuromorphic, however there is still an underlying misunderstanding of what NM really is and what it is capable of.NM has been a research topic in academia since the '80s, but only in the last decade this niche research field has widened to a much larger research community. From one side, this interest in NM has shed a new light on the potentials of Spiking Neural Networks (SNN), seen by many as the third generation of Neural Networks. From the other side, chip manufacturers have started to look at alternative computing 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 the 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, i.e. a clear 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 scepticism 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 a 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 fundamentally, there is a debate on best methods and practices 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 has 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.Therefore, in this Research Topic we welcome articles that aim to address the aforementioned issues related to practical application of SNN, specifically focusing on:• effective methodologies for unsupervised training of SNN on larger and more complex networks• 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