Advanced cognitive computing using adaptive spintronic materials INVITED

Abstract

The second-generation of Artificial intelligence (AI) has experienced intense growth due to its excellence in big-data applications. However, state-of-the-art AI lacks the advanced cognitive abilities seen in the brain, including the abilities to adapt to the environment, to understand context and to build coherent representations of the outside world. In addition, the implementation of second-generation AI with conventional CMOS uses large amounts of energy and area [1]. Spintronic nanodevices are particularly attractive for neuromorphic implementation of AI due to their small footprint, high endurance and low power consumption [2]. While these devices can implement ultra-low-power AI [3], they have yet to mimic the more rich, adaptive and computationally powerful behavior seen in the brain which forms the basis for advanced cognition. Here, we show the proof-of concept of a self-adaptive artificial neuron that utilizes adaptive spintronic materials that alter their structure and properties in response to external stimuli, in order to realize advanced cognitive abilities such as context-awareness and feature binding. The self-adaptive neuron is constructed from an artificial skyrmion lattice with five skyrmions hosted in a bilayer of thulium iron garnet (TmIG) and platinum (Pt). We report micromagnetic simulation results that show the neuron, when excited by an oscillating magnetic field, produces spin waves originating from skyrmion oscillations with a multi-frequency spectrum. The spectrum consists of four distinct resonant modes, identified as gyration, breathing, and hybridizations of both. Since each resonant frequency can represent one bit of information, the multi-frequent spectrum enables a large basis of information representation. Crucially, we show that both the amplitude and frequency of these resonant modes can be modulated by re-arranging the skyrmions in the lattice. This ability of the lattice to regulate its oscillatory dynamics in response to external input mimics a critical neural property called neuromodulation which plays a key role in advanced cognitive processes, including information transfer, decision-making, memory, object representation, consciousness etc. As a result of neuromodulation, the self-adaptive neuron demonstrates bio-plausible properties such as bursting and cross-frequency coupling [7], as well as advanced cognitive processes, namely context-awareness and feature binding. Context-awareness allows the neuron to make complex, high-level decisions while accounting for multiple factors to develop a deeper understanding of its circumstance. To implement context-awareness, the neuron receives a human spoken command and contextual information about the color of a box. Integrating both inputs, the neuron decides to open the box if and only if the human being gives the command to open and the box is not red in color. In addition, feature binding allows the neuron to correctly combine different segments of information (say, ‘red’ color and ‘circular’ shape) and construct a coherent percept (‘red circle’). To implement feature binding, a network of neurons receives visual information about two objects and processes information about their color and shape separately. The network then correctly binds together the two features (color and shape) to obtain coherent percepts of the original object. These results realize context-aware AI and the fusion of information from different sensory streams, which can have significant impact on human-machine collaboration, bio-medicine, smart energy, advanced manufacturing, agriculture and education. Adaptive materials can thus help realize neuromorphic circuits with advanced cognitive abilities for third generation AI, one that understands and adapts to a complex and ever-changing environment, learns without supervision from small datasets and better collaborates with human beings. **