Physics

Permanent URI for this collection

https://hdl.handle.net/10323/11897

Browse

Browsing Physics by Subject "Neuromorphic computing"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    Neuromorphic Computing with Antiferromagnetic Artificial Neurons
    (2024-01-01) Bradley, Hannah D; Tyberkevych, Vasyl; Tyberkevych, Vasyl; Slavin, Andrei; Khain, Evgeniy; Bozhko, Dmytro
    The increasing focus on Artificial Intelligence (AI) presents notable challenges, especially in the power requirements essential for training AI models. Consequently, there is a growing emphasis on neuromorphic computing, which aims to construct Artificial Neural Networks (ANNs) from artificial neurons to replicate the speed and efficiency of the human brain. With their significantly lower power consumption, spintronic devices acting as artificial neurons offer the potential for ANNs that rival conventional components. This innovative approach not only tackles the energy efficiency challenges but also paves the way for advancements in neuromorphic computing by integrating magnetic materials and spin-dependent effects. Antiferromagnetic (AFM) materials, characterized by their inherent THz frequencies, provide a unique opportunity for spintronic devices with ultra-fast dynamics. There's a proposal to utilize AFM materials to craft ultra-fast spin-Hall oscillators. These oscillators emit spiking signals resembling those of biological neurons, indicating their potential as artificial neurons. AFM oscillators exhibit exceptionally fast characteristics, including picosecond-scale spike widths and unique features absent in conventional artificial neuron models. This research examines the utilization of AFM oscillators as artificial neurons and their significance in the realm of neuromorphic computing. This dissertation begins with a comprehensive overview of relevant topics, covering the principles of spintronics, AFM materials, and neuromorphic computing. It examines the unique dynamics of AFM neurons, such as response latency and refraction time, which arise from an effective internal inertia. Then, it explores innovative AFM neuron circuits that exhibit functionalities unattainable by conventional artificial neurons, such as non-monotonic inhibition. Additionally, conventional learning algorithms like backpropagation are use to train AFM ANNs for pattern recognition. Finally, it leverages the similarities between AFM and biological neurons to model the biological neural network responsible for the withdrawal reflex. With their simplistic design and high speeds, AFM neurons exhibit energy efficiencies several orders of magnitude higher than traditional artificial neurons and even other spintronic designs. This suggests that AFM ANNs will excel at training AI models while addressing the energy crisis impeding technological progress. As a result, AFM neurons drive forward the progress of spintronic neuromorphic computing, providing a promising alternative for future technological development