IEEE Circuits and Systems Magazine - Q1 2020 - 58

On the one hand, this avoids the time continuity problem
described so far. At a generic time step k the input signal is
sampled, and each sample is multiplied by a j and accumulated to get y as in (12). The measurements have to be quantized and dispatched before the end of the time step. This
approach imposes time constraints tighter with respect to
a standard approach and due to the necessity to deal at the
same time with p different input signals (the time available
for all the required operations is Tw /n instead of Tw). However, at the end of the time windows all measurements have
already been converted and dispatched, and no additional
actions are required between the end of a time windows and
the beginning of the next one.
On the other hand, this approach makes the computation of y independent of the length of Tw . As observed
in [37], a Tw that is too long (that is actually a common
situation for biomedical signals such as ECG or EEG)
may lead to leakage problems in analog implementation
due to the discharge of the capacitors. In the light of
this, an architecture where performance of the integrator is independent of Tw is a sure advantage.
G. Dynamic CS Approaches
In many cases, it may be useful to tune some parameters
of a CS system on the particular input signal. For example, in biomedical signal acquisition, the dictionary
D used for reconstruction can be trained on the particular patient for improving performance. The training
phase, however, cannot be done by using compressed
measurements, but needs to be established before the
starting of the CS operating mode.
Such an approach is considered in [50]. The interesting aspect from the circuital point of view is that the designed system works with a feedback signal generated
by the reconstruction algorithm capable of enabling or
disabling the CS operating mode. Initially, the CS mode
is disabled (i.e., the device is working as a standard
Nyquist-rate converter) and the uncompressed input
EEG signal is sent to the receiver side which analyses it.
When enough data is collected to allow a good estimation of a trained dictionary, a signal is sent to the acquisition device to start the CS operating mode.8
A different case is taken into account in [52], where no
CS architectures are considered. Indeed, the design of a
low-power companion chip capable of detecting features
of the input signal is presented. The aim of the work is to
identify the type of the input signal among many possible
biosignals. In this way, it should be possible to adapt pa8
Note also that the trained dictionary characterizes the first of two
phases of the proposed decoding procedure. Here, the main signal
shape is approximately reconstructed by using a single atom from the
trained dictionary. In the second phase, signal details are recovered
using Daubechies wavelet transformation as sparsity basis.

IEEE CIRCUITS AND SYSTEMS MAGAZINE

rameters of the CS (for example, the number of measurements m, the time windows Tw, the number of samples
n per time window, and so on) to that of the input signal,
with advantages in terms of reconstruction quality and/
or energy required to sample the input signal.
VIII. Conclusion
CS is often thought of as a technique in which adaptation cannot play a significant role. Yet, a careful scan
of the literature reveals that many methods have been
developed to adjust either the encoding or the decoding
side of CS depending on the class of signals that has to
be acquired. Moreover, every hardware implementation
has to adapt the general paradigm to cope with realization constraints or resource budgets.
Our review shows that these designs are able to yield significant improvements that cannot be predicted by classical
theoretical guarantees dealing with worst-case analysis.
Yet, the possibility of substantially increasing the
compression ratio while obeying to the same requirements in terms of quality of the recovered signal, is a
key point to allow the effective introduction of CS in application in which resources must be carefully administrated. From this point of view, adapted CS is definitely
worth pursuing whenever designing the low-resources,
autonomous, ubiquitous sensing subsystems that will
be the backbone of future IoT applications.
Mauro Mangia (S'09-M'13) received
the B.Sc. and M.Sc. degrees in electronic engineering and the Ph.D. degree in
information technology from the University of Bologna, Bologna, Italy, in
2005, 2009, and 2013, respectively. He
was a Visiting Ph.D. Student with the Ecole Polytechnique Federale de Lausanne in 2009 and 2012. He is currently a Post-Doctoral Researcher with ARCES, Statistical Signal Processing Group, University of Bologna. His
research interests are in nonlinear systems, compressed
sensing, Internet of Things, ultra-wideband systems, and
systems biology. He was a recipient of the 2013 IEEE CAS
Society Guillemin-Cauer Award and of the 2019 IEEE BioCAS Transactions Best Paper Award. He received the
Best Student Paper Award at ISCAS2011. He was the Web
and Social Media Chair for ISCAS2018.
Fabio Pareschi (S'05-M'08-SM'19) received the Dr. Eng. degree (Hons.) in
electronic engineering from the University of Ferrara, Italy, in 2001, and the
Ph.D. degree in information technology
from the University of Bologna, Italy, in
2007, under the European Doctorate Project (EDITH). He
FIRST QUARTER 2020

IEEE Circuits and Systems Magazine - Q1 2020

Table of Contents for the Digital Edition of IEEE Circuits and Systems Magazine - Q1 2020