IEEE Circuits and Systems Magazine - Q2 2022 - 12

to achieve in SNN, due to the inherent asynchronous and
even driven nature of the SNN neurons. Because the input
and hidden state variables are represented by a sequence
of spike trains, special store-and-release neurons must be
used to gate and release the spike sequences in order to
synchronize them [60], [61]. It was not until recently that
some works presented techniques on translating recurrent
networks to SNNs [60], [61] or introducing dynamics
into the neuron and synapse models to implement a form
of recurrence [62]. We refer to these networks as recurrent
SNNs with synchronization.
A large set of neural networks use the recurrent structure
to stabilize signals and suppress noises. No synchro(a)
(b)
Input
Reservoir
(c)
Figure
5. ANN-inspired topologies.
Knowledge Links
Read Out
Symbols
nization among neurons is attempted in these networks.
The feedback and the input eventually reach a dynamic
equilibrium, which defines the network state. For example,
Echo State Network (ESN), also referred to as reservoir
network, is a variation of recurrent networks which
consists of a hidden layer (reservoir), containing neurons
randomly connected to each other with fixed weights,
and connected to the output layer which feeds back to
the reservoir with plastic weights [63]-[69] as shown in
Figure 5(c). In a reservoir network, the output layer is
used to classify the state of the network. Learning takes
place only in the output layer, which consists of conventional
(i.e. non-spiking) neurons. Another example is the
spiking confabulation network. Cogent confabulation is a
connection-based cognitive model that captures correlations
among features at the symbolic level, as shown in
Figure 6. It describes the basic dimensions of the observation
using a set of features referred to as lexicons. The attributes
of a given feature are referred to as the symbols,
which are analogous to neurons in the biological nervous
system. Their pairwise conditional probabilities are referred
to as the knowledge links. When implemented with
Bayesian spiking neurons as in [70], [71], the neurons interact
with each other and the equilibrium state of the
spiking rates infers the likelihood of the symbols represented
by each neuron.
A large number of SNNs are arranged to follow biological
neural architectures. For example, simultaneous
localization and mapping (SLAM) [72] networks get inspiration
from the navigation system in the hippocampus
and entorhinal cortex of rats, where different types
of spatially-tuned neurons were found: the head-direction
cells are sensitive to the heading direction of the animal,
place cells are active each time the animal visits a particular
part of the environment, and grid cells presumably
perform path integration [73]. As another example, Winner-Take-All
(WTA) networks containing recurrent connectivity
between inhibitory and excitatory neurons are
common models to explain decision-making and action
selection in the cortex [74], [75]. They are widely used for
unsupervised learning and feature selection in SNNs [76],
[77]. Figures 7(a) and 7(b) illustrate the structure of the
two aforementioned networks.
The hierarchies in the sensory cortex are of particuLexicons
Figure
6. Confabulation network.
12
IEEE CIRCUITS AND SYSTEMS MAGAZINE
lar interests to research in sensor signal processing. The
mammalian olfactory system contains three major hierarchical
levels including the epithelium where the stimulus
enters the nervous system, the olfactory bulb (OB) where
the first transformation happens, and the piriform cortex
(PC) which integrates and stores the information relevant
for odor recognition [78], [79]. Such hierarchical network
structures have been used for recognition and decisionmaking
tasks in SNNs [80]-[82]. Medial Superior Olive
SECOND QUARTER 2022
IEEE Circuits and Systems Magazine - Q2 2022

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