308
G. J. Lim et al.
Input received at the dendrites come from many other neurons, and can either be
excitatory or inhibitory. The inputs contribute to an analog voltage signal called the
membrane potential. The axon hillock operates on an all-or-none law that sums all the
input signals and triggers an action potential when a threshold potential is breached.
When sufficient excitatory input exceeds the potential threshold, an action potential
will propagate along the axon to the axon terminals. Aside from the analog and
digital mechanisms found within the brain, there are also dynamic, time-dependent
features. For example, in the spike-timing-dependent plasticity (STDP) process, the
synaptic strength between neurons are adjusted based on the relative timing between
pre- and post-synaptic activity. Neuron behaviour vary vastly depending on their type
and function [44]. In designing an artificial neuron/synapse, we trade-off between
device complexity, computational power efficiency, and bio-fidelity.
4 Spintronics for Neuromorphic Engineering
Most of the ongoing research work in artificial neural networks and machine learning
is through a top-down approach of developing algorithms to mimic biological neural
processes while still relying on hardware based on mature CMOS technologies.
Several takes on artificial neurons, synapses, and neural network hardware using
conventional trasistors include the IBM TrueNorth and Intel Loihi chips. Such chips
attempt to emulate biological neuronal and synaptic behaviour using transistors,
but not necessarily capture the same bio-fidelity of actual biological neurons and
synapses. However, properties and characteristics unique to emerging technologies
such as spintronics enable a bottom-up approach in the development of brain-inspired
computational hardware with greater bio-fidelity. In this section, we look at various
spintronic devices and discuss how they can be used to mimic behaviours, mechanisms, and features of the biological brain, as well as in building functional features,
efficient computational primitives, and circuits for real world applications.
4.1 Spintronic Synapses
The biological neuron receives excitatory and inhibitory post-synaptic potentials
along dendrites as shown in Fig. 10a. The summation of the input potentials can
ultimately lead towards generating an action potential. In an artificial neural network,
synapses with synaptic weights w m,n determine the connection strengths between the
pre- and post-synaptic neurons. The post-synaptic neuron N n receives inputs I m from
pre-synaptic neurons as shown in Fig. 10b, such that N n =
I m ·w m,n . An equivalent
circuit in the form of a synaptic crossbar array is shown in Fig. 10c. In the crossbar
array, a programmable device with memristive or analog-like behaviour is used to
encode the synaptic weight at each crosspoint. The output at each neuron is then the
sum of dot products between voltage inputs and synaptic conductance.
G. J. Lim et al.
Input received at the dendrites come from many other neurons, and can either be
excitatory or inhibitory. The inputs contribute to an analog voltage signal called the
membrane potential. The axon hillock operates on an all-or-none law that sums all the
input signals and triggers an action potential when a threshold potential is breached.
When sufficient excitatory input exceeds the potential threshold, an action potential
will propagate along the axon to the axon terminals. Aside from the analog and
digital mechanisms found within the brain, there are also dynamic, time-dependent
features. For example, in the spike-timing-dependent plasticity (STDP) process, the
synaptic strength between neurons are adjusted based on the relative timing between
pre- and post-synaptic activity. Neuron behaviour vary vastly depending on their type
and function [44]. In designing an artificial neuron/synapse, we trade-off between
device complexity, computational power efficiency, and bio-fidelity.
4 Spintronics for Neuromorphic Engineering
Most of the ongoing research work in artificial neural networks and machine learning
is through a top-down approach of developing algorithms to mimic biological neural
processes while still relying on hardware based on mature CMOS technologies.
Several takes on artificial neurons, synapses, and neural network hardware using
conventional trasistors include the IBM TrueNorth and Intel Loihi chips. Such chips
attempt to emulate biological neuronal and synaptic behaviour using transistors,
but not necessarily capture the same bio-fidelity of actual biological neurons and
synapses. However, properties and characteristics unique to emerging technologies
such as spintronics enable a bottom-up approach in the development of brain-inspired
computational hardware with greater bio-fidelity. In this section, we look at various
spintronic devices and discuss how they can be used to mimic behaviours, mechanisms, and features of the biological brain, as well as in building functional features,
efficient computational primitives, and circuits for real world applications.
4.1 Spintronic Synapses
The biological neuron receives excitatory and inhibitory post-synaptic potentials
along dendrites as shown in Fig. 10a. The summation of the input potentials can
ultimately lead towards generating an action potential. In an artificial neural network,
synapses with synaptic weights w m,n determine the connection strengths between the
pre- and post-synaptic neurons. The post-synaptic neuron N n receives inputs I m from
pre-synaptic neurons as shown in Fig. 10b, such that N n =
I m ·w m,n . An equivalent
circuit in the form of a synaptic crossbar array is shown in Fig. 10c. In the crossbar
array, a programmable device with memristive or analog-like behaviour is used to
encode the synaptic weight at each crosspoint. The output at each neuron is then the
sum of dot products between voltage inputs and synaptic conductance.