Spintronics for Neuromorphic Engineering
311
Fig. 12 a A non-step transfer function. b The resistance state of the DW neuron MTJ scales with
DW position
4.2.2 Non-step Neurons
The transfer function that models a neuronal output is typically non-stepwise such
as in Fig. 12a. In artificial neural networks, common transfer or activation functions
include the sigmoid, hyperbolic tangent, or more popularly, the rectified linear unit
(ReLU). To achieve beyond binary states in spintronic neuron devices, one would
require multi-domain formation in the free layer. Rather than existing as either one of
two possible magnetization states, a multi-domain free layer may form two oppositely
magnetized domains separated by a DW. The relative proportions between domains
can be adjusted by manipulating the DW position allowing for multiple resistance
states to be exhibited in a DW neuron MTJ [47, 50, 51]. The DW position can be
manipulated by current-induced SOT, and the resistance state of the device can be
read out using an MTJ [52]. The divider circuit in Fig. 12b operates such that an
increase in parallel orientation of the FL domain leads to lower resistance of the DW
neuron MTJ. This drives the output transistor to approximately match I out and I in
linearly.
4.2.3 (Leaky-)Integrate-and-Fire Neurons
The biological neuron receives multiple inputs from multiple unsynchronised presynaptic neurons. The excitatory and inhibitory inputs may not be received all at the
same time, but across a span of time. These multiple inputs trigger the neuron when
a threshold is reached and describes the IF functionality. Consider a similar structure
to the non-step neuron discussed earlier, with multiple short injected current pulses
to move the DW across the device, as depicted in Fig. 12a. At a certain point, the DW
would have crossed a threshold position such that the resistance state of the neuron
MTJ would cause the output inverter to be triggered as shown in Fig. 12b. The IF
functionality therefore describes the biological brain’s ability to summate all inputs
over time and trigger the neuron when the threshold potential is reached. However, it
does not address the potential decay as observed in biological systems, which further
adds time-dependence to the input stimulus.
311
Fig. 12 a A non-step transfer function. b The resistance state of the DW neuron MTJ scales with
DW position
4.2.2 Non-step Neurons
The transfer function that models a neuronal output is typically non-stepwise such
as in Fig. 12a. In artificial neural networks, common transfer or activation functions
include the sigmoid, hyperbolic tangent, or more popularly, the rectified linear unit
(ReLU). To achieve beyond binary states in spintronic neuron devices, one would
require multi-domain formation in the free layer. Rather than existing as either one of
two possible magnetization states, a multi-domain free layer may form two oppositely
magnetized domains separated by a DW. The relative proportions between domains
can be adjusted by manipulating the DW position allowing for multiple resistance
states to be exhibited in a DW neuron MTJ [47, 50, 51]. The DW position can be
manipulated by current-induced SOT, and the resistance state of the device can be
read out using an MTJ [52]. The divider circuit in Fig. 12b operates such that an
increase in parallel orientation of the FL domain leads to lower resistance of the DW
neuron MTJ. This drives the output transistor to approximately match I out and I in
linearly.
4.2.3 (Leaky-)Integrate-and-Fire Neurons
The biological neuron receives multiple inputs from multiple unsynchronised presynaptic neurons. The excitatory and inhibitory inputs may not be received all at the
same time, but across a span of time. These multiple inputs trigger the neuron when
a threshold is reached and describes the IF functionality. Consider a similar structure
to the non-step neuron discussed earlier, with multiple short injected current pulses
to move the DW across the device, as depicted in Fig. 12a. At a certain point, the DW
would have crossed a threshold position such that the resistance state of the neuron
MTJ would cause the output inverter to be triggered as shown in Fig. 12b. The IF
functionality therefore describes the biological brain’s ability to summate all inputs
over time and trigger the neuron when the threshold potential is reached. However, it
does not address the potential decay as observed in biological systems, which further
adds time-dependence to the input stimulus.