IEEE Circuits and Systems Magazine - Q2 2022 - 13

(MSO) in the mammalian auditory pathway is responsible
for sound localization. They are organized spatially as a
place map of location [83], [84] and at a higher level these
subgroups are subsequently also organized into frequency
selective clusters. These MSO systems inspired [85],
[86] to utilize SNNs for sound classification and localization.
Attractor networks, whose activity tend towards
dynamic stability, have been posited to help explain eye
control, working memory, head direction, locomotion and
olfaction [87] and thus have been utilized for such control
tasks in SNNs [88]-[90].
E. Learning
The ability of synaptic connections to change their efficacy
is referred to as synaptic plasticity. This is thought to be the
basic mechanism underlying learning and memory in biological
neural networks [91]. Various forms of synaptic plasticity
co-exist. Some are determined only by the history of
presynaptic stimulation, independently of the postsynaptic
responses [92]-[94]. Others depend on the temporal order
of pre- and postsynaptic activities [92], [95], [96].
In general, synaptic potentiation (i.e. the increase of
synaptic efficacy) is observed when presynaptic spikes
precede postsynaptic spikes, as it indicates a causal relationship.
The reversed order of spikes induces synaptic
depression (i.e. the decrease of synaptic efficacy). This
phenomenon is called Spike-Timing-Dependent-Plasticity
(STDP). It can be used for unsupervised learning. A popular
choice for the STDP rule [97] for potentiation wT +
depression wT - is given as:
TT
TT
wA t
wA t
++ +
-- -
=+
=exp
exp
^
^
-
+
/
/
x
x
h
h
for
for
T2
T1
t
t
and
(11)
where
x+ and x- determine the ranges of pre- to postsynaptic
interspike intervals over which synaptic strengthening
and weakening occur. A+
and Adetermine
the
maximum amounts of synaptic modification, and tT is
the time of the postsynaptic spike minus the time of the
presynaptic spike.
A wide range of SNN algorithms that can learn temporal
spike patterns employ more biologically realistic LIF
neuron models with alpha or dual-exponential synapse
[57], [77], [98]-[100] as shown in Section II-B. In these
models, the PSP decays exponentially over time, hence,
can be utilized as a metric to reflect temporal dependency.
This type of neuron does not simply accumulate
weighted spikes as membrane potential, instead, it integrates
weighted time varying PSP, hence exhibiting complex
temporal dynamic behavior. Various STDP learning
rules have been proposed that update the weight based
on different traces. Traces are decaying state variables reflecting
the temporal history of input and output spikes.
They correspond to the ion concentrations in a biological
neuron. The PSP is an example of a trace variable. A pairwise
STDP rule using traces [101] is given as:
Environment
Agent
Egocentric
Observation
Sensors
Distance
(Sonar,
Lidar, etc)
Error Correction
IEEE Circuits and Systems Magazine - Q2 2022

Table of Contents for the Digital Edition of IEEE Circuits and Systems Magazine - Q2 2022
