IEEE Circuits and Systems Magazine - Q1 2022 - 18

common ANN topologies and their potential application
to CR systems. Those called generic ANNs-where synaptic
weights are trained by using examples-can either
be feedforward or recurrent, and are usually associated
to applications like artificial vision or voice recognition,
where logic programming presents some limitations. The
so-called Deep Neural Network (DNN) has at least more
than one hidden layer of neurons and it operates in a feedforward
way. These ANNs use costly offline training methods
which require huge amount of datasets. They also
present very good performance for some tasks like face
recognition, movement detection, etc. Another topology
of ANNs-shown in Fig. 8-is the so-called Spiking Neural
Network (SNN) [43]. In this type of ANNs, neurons update
their state only when they receive an input spike-in a
similar way as it happens in biology-and there is a temporal
correlation between spikes, which becomes crucial
for the correct modeling of these kinds of ANNs [41], [42].
There are many other classification criteria of ANNs.
One of then is the degree of connectivity of their neurons,
so that ANNs can be divided in fully-connected ANNs-
where each neuron in a given layer is connected to
all neurons of the remaining layers in a feed-forward
way-or Convolutional Neural Networks (CNNs), where
each neuron is connected to a given subgroup of neighboring
layers in a projective/receptive field (see Fig. 7). Another
classification criterion deals with the way ANNs
are trained, thus dividing ANNs into supervised-learning,
unsupervised-learning or self-learning, depending
on the data provided to the ANN to get them trained.
Finally, another criterion deals with the implementation
itself, which can be either in software-by using
many diverse language programs such as Python, C,
MATLAB, etc.-or in hardware, either in FPGAs or embedded
in a chip.
B. Neuromorphic Computing
One of the ANN approaches that is gaining more and more
relevance in the last years is the so-called neuromorphic
computing, which is inspired by the information processing
of the human brain [44], [45]. This is the case
of the TrueNorth processor, designed by IBM in 28 nm
process [46]-with a high energy efficiency of 26 pJ
per synapse-making it viable for a broad spectrum of
commercial applications. Another milestone is the Loihi
chip, developed by Intel in a 14 nm technology, with the
capacity to emulate 130,000 artificial neurons [47]. Neural
computing is also beginning to be implemented in
modules embedded in conventional processors to perform
very diverse ML functions. For example, the A1X
Bionic series processors-developed by Apple for their
mobile devices and more recently in their M1X series
processors included in recent generations of their computers
18
IEEE CIRCUITS AND SYSTEMS MAGAZINE
FIRST QUARTER 2022
Memristors
Spike Timing Dependent Plasticity
IT-1R
Ox-RAM
TiN
TiN
HfO2
Ti
AI Engine
Neurons
-
+
Vref
Memristor-Based (4 × 4) Crossbars
Spiking
Circuit
Synapsis in Nature
Figure 8. Illustration of spiking neural networks based on memristive-based crossbars and their application to neuromorphic processors for AI-managed CR.
Rx
RF
DSP and
Neural
Processor
Tx
RF
ASP
ADC
DAC
ASP

IEEE Circuits and Systems Magazine - Q1 2022

Table of Contents for the Digital Edition of IEEE Circuits and Systems Magazine - Q1 2022