Shape-Based Activation Functions in Magnetic Domain Wall Leaky Integrate-and-Fire Neurons for Artificial Intelligence Applications

Although standard von-Neumann architectures are very well suited for the processing of highly structured data, various aspects of these systems – such as non-volatility – make them less practical for processing unstructured, real-world data. Therefore, it is desirable to mimic the human brain in order to provide significant improvements in the computation of such data. We previously proposed three biomimetic leaky integrate-and-fire (LIF) neurons that intrinsically provide all three neuronal functionalities without the use of any external circuitry [1]-[3], which in turn provide improvements in terms of area overhead and energy consumption compared to previous LIF neurons. However, it is desirable to further improve the biomimicry of these neurons by implementing certain mathematical functions during device operation. By altering the shape of the neurons, we can implement various leaking characteristics, including the linear and sigmoidal leaking characteristics we will discuss in this work.

[1] Hassan, et al., JAP, 2018.

[2] Brigner, et al., JxCDC, 2019.

[3] Brigner, et al., TED, 2019.