Modern communication networks are rapidly evolving into sophisticated systems combining communication and computing capabilities. Computation at the network edge is the key to supporting many emerging applications, from extended reality to smart health, smart cities, smart factories and autonomous driving. SONATA is motivated by the fact that the large scale adoption of edge intelligence technology, while benefiting human productivity and efficiency, will result in a surge of data and computation in mobile networks, which, in turn, will exacerbate their already significant energy consumption. SONATA is an interdisciplinary effort to tame this growing energy demand by combining memristive hardware and energy harvesting technologies with novel machine learning algorithms and physical layer communication techniques. In particular, we want to combine the energy efficient in-memory computing and learning potential of memristive devices with an “over-the-air computation (OAC)” approach to edge learning, which turns the air from a purely communication medium to a computation unit. Our project not only aims at reducing the energy requirements of edge learning systems drastically, but also focuses on making them robust against stochastic failures, due to unreliable hardware or energy sources. We will exploit tools from circuit design, coding theory, wireless communications, machine learning and network science to achieve these goals. Results from SONATA will open up new directions for research and development of technologies that will allow mobile systems to offer the much anticipated communication and computing services in a sustainable manner.

In recent years, we have witnessed an explosion of artificial intelligence (AI) applications which will continue to grow over the next decade. An intelligent and digitized society will be ubiquitous, enabled by increased advances in nanoelectronics. Key drivers will be sensors interfacing with the physical world and taking appropriate action in a timely manner while operating with energy efficiency and flexibility to adapt. The vast majority of sensors receive analog inputs from the real world and generate analog signals to be processed. However, digitizing these signals not only creates enormous amount of raw data but also require a lot of memory and high-power consumption. As the number of sensor-based IoTs grows, bandwidth limitations make it difficult to send everything back to a cloud rapidly enough for real-time processing and decision-making, especially for delay-sensitive applications such as driverless vehicles, robotics, or industrial manufacturing. In this context, PHASTRAC proposes to develop a novel analog-to-information neuromorphic computing paradigm based on oscillatory neural networks (ONNs). We propose a first-of-its-kind and novel analog ONN computing architecture to seamlessly interface with sensors and process their analog data without any analog-to-digital conversion. ONNs are biologically inspired neuromorphic computing architecture, where neuron oscillatory behavior will be developed by innovative phase change VO2 material coupled with synapses to be developed by bilayer Mo/HfO2 RRAM devices. PHASTRAC will address key issues 1) novel devices for implementing ONN architecture, 2) novel ONN architecture to allow analog sensor data processing, and 3) processing the data efficiently to take appropriate action. This “sensing-to-action” computing approach based on ONN technology will allow energy efficiency improvement 100x-1000x and establish a novel analog computing paradigm for improved future human-machine interactions.

Artificial neural networks represent a key component of neuro-inspired computing for non-Boolean computational tasks. They emulate the brain by using nonlinear elements acting as neurons that are interconnected through artificial synapses. However, such physical implementations face two major challenges. First, interconnectivity is often constrained because of limits in lithography techniques and circuit architecture design; connections are limited to 100s, compared with 10000s in the human brain. Second, changing the weight of these individual interconnects dynamically requires additional memory elements attached to these links. Here, we propose an innovative architecture to circumvent these issues. It is based on the idea that dynamical hyperconnectivity can be implemented not in real space but in reciprocal or k-space. To demonstrate this novel approach we have selected ferromagnetic nanostructures in which populations of spin waves – the elementary excitations – play the role of neurons. The key feature of magnetization dynamics is its strong nonlinearity, which, when coupled with external stimuli like applied fields and currents, translates into two useful features: (i) nonlinear interactions through exchange and dipole-dipole interactions couple potentially all spin wave modes together, thereby creating high connectivity; (ii) the strength of the coupling depends on the population of each k mode, thereby allowing for synaptic weights to be modified dynamically. The breakthrough concept here is that real-space interconnections are not necessary to achieve hyper-connectivity or reconfigurable synaptic weights. The final goal is to provide a proof-of-concept of a k-space neural network based on interacting spin waves in low-loss materials such as yttrium iron garnet (YIG). The relevant spin wave eigenmodes are in the GHz range and can be accessed by microwave fields and spin-orbit torques to achieve k-space Neural computation with magnEtic exciTations.