An Attention Mechanism-Based Adaptive Feedback Computing Component by Neuromorphic Ion Gated MoS2 Transistors

Neuromorphic computing is expected to bridge cognitive behaviors with computing systems in an efficient, expandable, and biologically inspired way. A pivotal cognitive behavior is the attention mechanism, which is highly important in filtering and regulating spatio-temporal information. Emerging neuromorphic devices hold prospect in utilizing their internal physical mechanisms for dynamic computing resources. Here, a basic top-down attention computing component consisting of a synaptic transistor and a neuron is proposed, where efficient information processing is realized by combining the inherent device dynamics and the feedback loop. A theoretical model is established in simulation to demonstrate the capabilities of such a computing system in information filtering and control. Notably, new dynamic circuit behaviors, such as conductance oscillation and activate function switching, are discovered from appropriate time parameters. The attention computing component contains rich dynamic behaviors, providing a power and area-saving method to construct high-complexity neuromorphic systems for spatio-temporal signal preprocessing and control.

Automated summary

Neuromorphic computing aims to connect cognitive behaviors with efficient computing systems in a biologically inspired way. Attention mechanism, an important cognitive behavior for filtering and regulating spatio-temporal information, can be efficiently processed using the dynamic capabilities of emerging neuromorphic devices. A basic top-down attention computing component comprising a synaptic transistor and a neuron is proposed and demonstrated to effectively filter and control information. The component exhibits new dynamic circuit behaviors, such as conductance oscillation and activate function switching, offering a power and area-saving method for constructing complex neuromorphic systems.