Content uploaded by Marcello Zanghieri
Author content
All content in this area was uploaded by Marcello Zanghieri on Jun 21, 2024
Content may be subject to copyright.
This is the final peer-reviewed accepted manuscript of:
M. Zanghieri, P. M. Rapa, M. Orlandi, E. Donati, L. Benini, S. Benatti,
“sEMG-driven Hand Dynamics Estimation with Incremental Online
Learning on a Parallel Ultra-Low-Power Microcontroller,” IEEE
TBioCAS 2024
The final published version is available online at:
https://ieeexplore.ieee.org/document/10559752
Rights/License:
The terms and conditions for the reuse of this version of the manuscript
are specified in the publishing policy. For all terms of use and more
information see the publisher’s website.
GENERIC COLORIZED JOURNAL, VOL. XX, NO. XX, XXXX 2017 1
sEMG-driven Hand Dynamics Estimation with
Incremental Online Learning on a Parallel
Ultra-Low-Power Microcontroller
Marcello Zanghieri,* Graduate Student Member, IEEE, Pierangelo Maria Rapa, Graduate Student
Member, IEEE, Mattia Orlandi, Graduate Student Member, IEEE, Elisa Donati, Member, IEEE, Luca
Benini, Fellow, IEEE , Simone Benatti, Member, IEEE
Abstract—Surface electromyography (sEMG) is a State-
of-the-Art (SoA) sensing modality for non-invasive human-
machine interfaces for consumer, industrial, and rehabilita-
tion use cases. The main limitation of the current sEMG-
driven control policies is the sEMG’s inherent variability,
especially cross-session due to sensor repositioning; this
limits the generalization of the Machine/Deep Learning
(ML/DL) in charge of the signal-to-command mapping. The
other hot front on the ML/DL side of sEMG-driven control is
the shift from the classification of fixed hand positions to
the regression of hand kinematics and dynamics, promis-
ing a more versatile and fluid control. We present an incre-
mental online-training strategy for sEMG-based estimation
of simultaneous multi-finger forces, using a small Temporal
Convolutional Network suitable for embedded learning-on-
device. We validate our method on the HYSER dataset,
cross-day. Our incremental online training reaches a cross-
day Mean Absolute Error (MAE) of (9.58 ± 3.89)% of the Max-
imum Voluntary Contraction on HYSER’s RANDOM dataset
of improvised, non-predefined force sequences, which is
the most challenging and closest to real scenarios. This
MAE is on par with an accuracy-oriented, non-embeddable
offline training exploiting more epochs. Further, we demon-
strate that our online training approach can be deployed
on the GAP9 ultra-low power microcontroller, obtaining a
Submitted on 22 February 2024.
This manuscript is submitted as an invited extension of
the conference paper M. Zanghieri et al., “Online unsupervised
arm posture adaptation for sEMG-based gesture recognition on
a parallel ultra-low-power microcontroller,” in 2023 IEEE Biomed-
ical Circuits and Systems Conference (BioCAS), 2023, pp. 1-5.
DOI: 10.1109/BioCAS58349.2023.10388902. IEEE Xplore URL:
https://ieeexplore.ieee.org/document/10388902.
*Corresponding author.
M. Zanghieri, P. M. Rapa, M. Orlandi, L. Benini, and S. Be-
natti are with the Department of Electrical, Electronic, and Informa-
tion Engineering, University of Bologna, 40136 Bologna, IT (e-mail:
{marcello.zanghieri2, pierangelomaria.rapa, mattia.orlandi, luca.benini,
simone.benatti}@unibo.it).
E. Donati is with the Institute of Neuroinformatics, University of Z¨
urich
and ETH Z¨
urich, 8057 Z ¨
urich, CH (e-mail: elisa@ini.uzh.ch).
L. Benini is also with the Integrated Systems Laboratory, Department
of Information Technology and Electrical Engineering, ETH Z¨
urich, 8092
Z¨
urich, CH (e-mail: lbenini@iis.ee.ethz.ch).
P. M. Rapa and S. Benatti are also with the Department of Engineer-
ing, University of Modena and Reggio Emilia, 41125 Modena, IT (e-mail:
pierangelomaria.rapa@unimore.it, simone.benatti@unimore.it).
This research was supported in part by the EU Horizon Europe
project IntelliMan (g.a. 101070136) and by the ETH Z¨
urich’s Future
Computing Laboratory funded by a donation from Huawei Technologies.
latency of 1.49 ms and an energy draw of just 40.4 uJ per
forward-backward-update step. These results show that our
solution fits the requirements for accurate and real-time
incremental training-on-device.
Index Terms—Continual Learning, Deep Learning, Elec-
tromyography, Embedded, Human-Machine Interaction,
Human-Machine Interfaces, Incremental Learning, Low-
Power, Machine Learning, Microcontroller, On-Device
Learning, Online Learning, Parallel Computing, Prosthetics,
Ultra-Low Power, PULP, Real Time, Regression, Temporal
Convolutional Networks, TinyML, Training-on-Device.
I. INTRODUCTION
Decoding surface electromyographic (sEMG) [1], [2] sig-
nals is nowadays a widespread approach for driving Human-
Machine Interfaces (HMIs) in an intuitive and non-invasive
way in scenarios involving robotics, prosthetics, and other
industry and consumer use cases [3]–[8]. The current dominant
approach exploits classical Machine Learning (ML) [9]–[15]
or Deep Learning (DL) [16]–[20] to automatize and optimize
the signal-to-command strategy, mapping the sEMG to hand
positions, movements, or forces.
The main obstacle to the ML/DL ability to generalize
accurately and robustly is represented by the numerous vari-
ability factors inherent in the sEMG signal, such as anatomical
variability, fatigue, skin perspiration, and electrode reposi-
tioning. The current SoA ML/DL methodology has tackled
the variability issue by using statistical methods to align
data across days [21] or by producing larger datasets and
performing multi-session training of the models [6], [22]–[24],
i.e., by training deep models on sEMG data acquired in diverse
conditions, postures, electrode placements, or users [10], [25].
This approach is mainly applied to DL since multi-session
training benefits Deep Neural Networks (DNNs) more than
non-deep ML [26], [27]. However, multi-session training has
high computational and memory costs since it requires stor-
ing training datasets representative of several conditions and
involves iteratively training deep models for several epochs.
In addition, collecting sEMG datasets sufficiently large for
multi-session training remains challenging since curating and
releasing high-quality datasets requires several participants and
sessions and is labor-intensive [6].
2 GENERIC COLORIZED JOURNAL, VOL. XX, NO. XX, XXXX 2017
Another major limitation of the ML/DL approach to sEMG-
driven HMIs is the focus on classification, i.e., the recognition
of hand postures or gestures. Even if accurate, sEMG recog-
nition limits the control to a discrete set of fixed predefined
movements. In contrast, regression of hand kinematics or
dynamics promises a more fluid and versatile control. Recent
advances in sEMG regression have been made thanks to
public sEMG datasets [28]–[30] and accurate models [31]–
[33]. However, very few sEMG regression works target the
deployment and profiling of the regression algorithms onto
embedded computational devices [34], [35].
To advance wearable sEMG-driven HMIs, it is necessary to
put effort into the hardware-software co-design required to port
the accurate SoA algorithms into the domain of Tiny Machine
Learning (TinyML) [36]–[38], i.e., the approach to ML that
targets resource-constrained platforms and seeks to optimize
the tradeoff between accuracy and execution requirements such
as memory footprint, computational burden, and latency [39].
In this work, we propose a strategy of incremental online
learning for simultaneous multi-finger force estimation on an
embedded platform in real time. The method is based on re-
training a small Temporal Convolutional Network (TCN) [40],
[41], designed for combining accuracy and a reduced compu-
tational and memory budget, for time-domain processing of
raw sEMG [24], [34], [42] and other raw biosignals [7], [43],
[44]. Our contribution is three-fold:
•we propose an incremental online learning setup to learn
new multi-finger force patterns processing the retraining
data in streaming mode, consuming them one-by-one, and
seeing them just once as in an online adaptation session
on an embedded device that does not save the whole
retraining set into volatile or non-volatile memory;
•we validate our method on the HYSER sEMG regression
dataset [29], obtaining a cross-day Mean Absolute Er-
ror (MAE) of (9.58±3.89)% of the Maximum Voluntary
Contraction (MVC) on HYSER RANDOM, the dataset’s
most challenging section containing improvised, non-
predefined force sequences, closest to real scenarios; this
MAE is on par with a baseline offline training exploiting
more computational resources, and compatible with the
literature addressing HYSER in easier settings;
•we deploy and profile the retraining on the GAP9 parallel
ultra-low power Microcontroller Unit (MCU), obtaining
a power draw of 27.1 mW; for each forward-backward-
update step, we measure an energy consumption of
40.4µJand a latency of 1.49 ms, shorter than the input
data frequency used for training and thus consistent with
the requirements of real-time training-on-device.
The present work is an extension of our previous paper [25]
and targets the same challenge of sEMG’s pattern variability
across different conditions and sessions. As schematized in
Table I, the present work contributes an incremental re-
learning strategy which is more powerful than [25] since it
advances from sparse, low-count to high-density, high-count
electrodes, from classification of fixed gestures to regression,
and from cross-arm-posture adaptation to incremental learning
of new simultaneous multi-finger force patterns. This progress
is promising for the future of intuitive and reliable non-invasive
wearable HMIs.
To allow the research community to reproduce our work,
we make the code developed for this paper available open-
source.1
II. RE LATE D WORK
This section outlines the SoA of the ML/DL-based HMI
control policies that address the two main current challenges
of sEMG-driven HMIs: the inherent variability of the sEMG
signal patterns (II-A) and the advance from recognition to
regression (II-B).
A. sEMG Inherent Variability
In the ML/DL framework for sEMG-driven control, the SoA
methods for addressing the sEMG signal’s inherent variabil-
ity are mainly multi-session training and model adaptation.
In multi-session training, a ML/DL model is refit on data
collected in different conditions, e.g., diverse users, sessions
(involving a different positioning of the sEMG sensors), or
arm postures [10]. In contrast, model adaptation only involves
partial retraining on data from a novel sEMG session. Multi-
session training and model adaptation are mainly applied to
DL models since multi-session training benefits deep models
more than non-deep ML [26], and the deep nets’ modularity
allows to only retune a subset of layers or modules, e.g.,
batch-normalizations in AdaBN [45], [46], or a model’s final
classifier block in continual learning [27].
Multi-session training and deep net adaptation improve
recognition accuracy but imply a large computational burden
and a high memory footprint. On the one hand, multi-session
training requires storing multiple sEMG sessions as training
sets to be merged and training the model on a server. On
the other hand, adaptation based on fine-tuning can lower the
memory footprint. For instance, latent replay stratagems [47]
only save a reduced subset of intermediate activations [27].
Other solutions only retune the final or BN layers, avoiding
the computational load of back-propagation [46]. However,
these methods were not deployed on resource-constrained
computing devices.
To the best of our knowledge, we are the first to propose an
on-device learning setup for sEMG regression, which fits the
resource constraints of embedded devices suitable for wearable
HMIs.
B. sEMG Regression
The estimation of the hand kinematics and dynamics,
framed as a ML/DL regression task, promises a more fluid
and natural control compared to the classification that maps
the sEMG to a limited, fixed set of static positions.
The literature addressing sEMG regression can be classified
based on the variable chosen as the estimation target: kinemat-
ics variables (e.g., joint angle and joint velocity) or dynamic
variables (finger forces). Joint angles constitute the most rep-
resentative variable for hand kinematics and are the quantities
1https://github.com/pulp-bio/incremental_hyser/
M. ZANGHIERI et al.: PREPARATION OF PAPERS FOR IEEE TRANSACTIONS AND JOURNALS (FEBRUARY 2017) 3
TABLE I: Outline of the present work as an extension of our previous publication [25].
WORK sEMG sensing ML task Addressed variability Validation
Zanghieri et al. [25]
(extended here)
4channels:
low-density, low-count
classification:
fixed gestures new arm postures within-session
This work 64 channels:
high-density, high-count
regression: simultaneous
multi-finger forces new force patterns cross-day
most used as a target variable. For instance, [28] targets the
estimation of the joint angles of a dataglove, using a Wiener
filter and attaining a median R2of 0.63. However, this work
does not investigate other algorithms to improve the regression
score since it focuses on the control quality perceived by users.
The work [31] employs a Long Short-Term Memory deep
neural network on the same dataset as [28], reaching a Mean
Absolute Error (MAE) of 7.04◦. The limitation of [31] is that
it does not explore convolutional networks, which are more
amenable to parallelization and hence more suitable for em-
bedded control. An alternative kinematic approach targets joint
velocity to model finger motion in addition to finger position.
For example, the work [15] estimates velocity using a hybrid
classification-regression setup, which is based on thresholding
the speed’s dynamic range into 3intervals; though novel, this
method still limits the estimation to the recognition of discrete
classes.
A different and independent research direction in sEMG
regression aims to model hand dynamics, i.e., finger forces.
The work [48] targeted a multiple-Degrees-of-Freedom (multi-
DoF) force estimation but is limited to grasp movements. In
contrast, in the present work, we address the richer set of
force patterns present in the public sEMG HYSER dataset
(III-A) [29].
The limitation of all the sEMG regression works mentioned
above is that they do not focus on the embedded deployment
onto resource-limited computational platforms. The interest
for accuracy alone is common to other sEMG regression
works that implement convolutive deep learning [49], [50].
The deployment of a SoA TCN for sEMG hand kinematics
regression is addressed in [34], matching the SoA accuracy on
the NinaPro DB8 [28] dataset and deploying the TCN onto the
parallel ultra-low power microcontroller GAP8 (predecessor of
GAP9, which we use in this work). However, [34] does not
address cross-session validation and online learning, which we
target in the present work.
III. MATERIALS & METHODS
A. HYSER Dataset
This work targets the High-densitY Surface Electromyo-
gram Recordings (HYSER)2[29], a rich sEMG dataset re-
leased for research on sEMG recognition and force regres-
sion. HYSER features 20 healthy participants undergoing 2
acquisitions in separate days at a distance of 3to 25 days
(8.5±6.7days on average). sEMG data are collected with
four 8×8HD-sEMG arrays (256 electrodes in total) on the
2https://www.physionet.org/content/hd-semg/2.0.0/
forearm covering the extensor and flexor muscles, employ-
ing an OT Bioelettronica Quattrocento system sampling at
2048 samples/s. Force ground-truth signals are collected in
isometric contractions with a sensor-amplifier pair for each
finger, employing Huatran SAS sensors and Huatran HSGA
amplifiers sampling at 100 samples/s. In all the experiments
of this work, we comply with the recommended filtering of
the HYSER data [29]: (i) on the HD-sEMG data, an 8th-order
band-pass Butterworth filter with pass-band between 10 Hz
and 500 Hz, and a notch filter to attenuate the power line
interference at 50 Hz and harmonics up to 400 Hz; (ii) on the
force data, an 8th-order lowpass Butterworth filter with cut
frequency 10 Hz.
The HYSER dataset is composed of 5sub-datasets:
1) PR: pattern recognition on 34 hand gestures (not used
in this work);
2) MVC: trials for determining the MVC of every finger’s
flexion and extension;
3) 1-DoF: single-finger contractions for 1-DoF force esti-
mation, subdivided into 5trials ×5fingers;
4) 5-DoF: multi-finger contraction following prescribed
combinations and trajectories for 5-DoF force estimation
in controlled conditions, subdivided into 5trials ×5
fixed finger combinations;
5) RANDOM: multi-finger contractions performed in a fash-
ion defined random task, i.e. with no prescribed protocol
of combinations or trajectories, subdivided into 5trials
(each lasting 25 s, performed with a 5 s inter-trial rest to
prevent muscle fatigue);
This work uses datasets 2to 5, oriented to simultaneous multi-
finger force estimation. In particular, HYSER RANDOM is the
closest to real conditions in that the force patterns and dynamic
ranges are not fixed and are allowed to differ between Day 1
and Day 2, making HYSER RANDOM the most challenging
part of HYSER.
Most works that have addressed the HYSER dataset have
targeted discrete gesture recognition on the HYSER’s PR sub-
dataset; only little research has been conducted on continuous
force estimation on HYSER’s 1-DoF,5-DoF, and RANDOM
sub-datasets. We report the SoA literature on HYSER regres-
sion in Table II, also comparing it against this work. It is
worth remarking that this work addresses the two HYSER’s
most challenging settings: (i) cross-day validation, and (ii)
the RANDOM dataset, with force patterns and dynamic ranges
that are not fixed by protocol and allowed to differ between
Day 1and Day 2. Targeting these two scenarios means using
the HYSER dataset in the way closest to real, out-of-the-lab
conditions. We provide an in-depth discussion in Section V.
4 GENERIC COLORIZED JOURNAL, VOL. XX, NO. XX, XXXX 2017
TABLE II: Literature summary of force estimation on the HYSER HD-sEMG dataset.
WORK Purpose Algorithm HYSER
sub-dataset Validation Results HW?
incremen-
tal/online
learning?
Jiang et al.
[29] (2021)
HYSER
release
FIR
kernel RANDOM within-day,
leave-1-trial-out RMSE = (8.57 ±5.27)% MVC ✗ ✗✗
Jiang et al.
[51] (2022)
channel
selection
FIR kernel,
random masks 5-DoF cross-day,
leave-1-subject-out RMSE = (8.66 ±0.96)% MVC ✗ ✗✗
Jiang et al.
[52] (2023)
robustness,
physiological
explainability
deep forests 1-DoF
cross-day:
train on Day 1,
test on Day 2
RMSE = (8.0±2.3)% MVC
rPearson = 0.900 ±0.101
R2= 0.631 ±0.172
✗ ✗✗
Wu et al.
[53] (2023)
motor units
extraction
gCKC BSS,
cumulative
spike train
1-DoF no ML-style
validation rPearson = 0.908 ±n.a. ✗ ✗✗
This work
incremental online
learning embedded
on a parallel
ultra-low-power MCU
TCN RANDOM
cross-day:
train on Day 1,
test on Day 2
MAE = (9.58 ±3.89)% MVC ✓ ✓✓
= taken
= not taken
Fig. 1: Uniform downsampling applied to the HYSER HD-
sEMG channels. The same 2×downsampling in both direc-
tions is applied to all 4HYSER’s 8×8sensor patches, lowering
the channel count from 256 to 64.
B. Tiny Temporal Convolutional Network
We address force estimation by designing an extremely
compact Temporal Convolutional Network (TCN) [40]–[42]
that fits the resource constraints of embedded computational
platforms in the spirit of TinyML. The TCN’s complete
structure is detailed in Table III. The net is composed of
5convolutional blocks followed by a final 3-layer fully con-
nected block; all convolutions and poolings are 1-dimensional,
acting only in the time dimension, and have kernel 2, stride
2, and no padding (also termed valid padding).
The model’s input is constituted by sEMG time-windows of
size 64 channels ×63 samples, corresponding to a duration of
∆ = 30.8 ms; these windows are taken with slide 64 samples,
i.e., 31.2 ms, in both training and validation. The 64 chan-
nels are obtained by undersampling HYSER’s original 256
channels uniformly over the 4sensor patches; each of the 4
HYSER’s 8×8patches (III-A) is downsampled spatially by
2×along both dimensions, as illustrated in Fig. 1, yielding a
total downsampling factor of 4.
The model’s output is the 5-dimensional output y∈R5
HYSER regression datasets
Day 1
Day 2
1-DoF 5-DoF RANDOM
initialization training retraining retraining
cross-day
validations
Fig. 2: Scheme of the incremental protocol implemented on
the HYSER dataset.
of the last fully connected layer (without any non-linearity),
constituting the 5-finger forces estimates. In compliance with
the real-time TCN paradigm [40], [41] (which has been
yielding advances in the efficient processing of sEMG [24],
[34] and other raw biosignals [43], [44]), we adopt a causal
setup by pairing each ground-truth force yat time tywith
the sEMG window t∈[ty−∆, ty], so that each TCN pass
only takes into account data from the past, without any leak of
information from the future. In contrast, future data leak would
happen if pairing the ground truth with the sEMG signal in
t∈ty−1
2∆, ty+1
2∆.
This TCN is designed to be as compact as possible while ex-
ploiting the numerous sEMG sensors available on the dataset.
We do so by reducing the channels from 64 to 16 in the first
convolutional layer. As shown in Table III, the overall size of
this TCN is 3317 parameters and 150 kMAC, which is very
hardware-friendly for embedded devices [36]–[38].
C. Incremental & Online Learning Protocol
On top of the HYSER dataset, we define a training protocol
that is both incremental and online.
1) Incremental Learning:We employ the HYSER 1-Dof,
5-DoF, and RANDOM datasets (III-A) as a sequence of in-
creasingly complex patterns to be learned. The incremental
protocol (shwon in Fig. 2) follows three stages, performed
separately for each of HYSER’s 20 subjects:
M. ZANGHIERI et al.: PREPARATION OF PAPERS FOR IEEE TRANSACTIONS AND JOURNALS (FEBRUARY 2017) 5
TABLE III: Structure of the proposed tiny TCN.
BLOCK Layer Size # MAC (forward pass)
input parameters
(weights + biases) output
Convolutional 0 conv-ReLU 64 ×63 2064 16 ×62 127968 (86.01%)
max-pooling 16 ×62 16 ×31
Convolutional 1 conv-ReLU 16 ×31 528 16 ×30 15840 (10.65%)
max-pooling 16 ×30 16 ×15
Convolutional 2 conv-ReLU 16 ×15 264 8 ×14 3696 (2.48%)
max-pooling 8×14 8 ×7
Convolutional 3 conv-ReLU 8×7 136 8 ×6 816 (0.55%)
max-pooling 8×6 8 ×3
Convolutional 4 conv-ReLU 8×3 136 8 ×2 272 (0.18%)
max-pooling 8×2 8 ×1
(flatten) 8×1 8
Dense
dense-ReLU 8 72 8 72 (0.05%)
dense-ReLU 8 72 8 72 (0.05%)
dense 8 45 5 45 (0.03%)
TOTAL Total parameters: 3317 Total MAC: 149 ·103
•Stage 0:we initialize the TCN (III-B) and train it
on HYSER 1-DoF, Day 1, to learn single-finger force
estimation.
•Stage 1:we retrain the net on HYSER 5-DoF, Day 1, to
extend the learning to fixed combinations of multi-finger
forces;
•Stage 2:we retrain on HYSER RANDOM, Day 1, to teach
the model more diverse, improvised multi-fingers forces
not belonging to predefined finger combinations.
All (re)trainings only see data from the Day 1of the HYSER
sub-dataset that is the stage’s target. After every (re)training,
we validate on Day 2of 1-Dof,5-DoF, and RANDOM
to compare the regression quality attained at each stage; in
particular, we validate on Day 1of 5-DoF and RANDOM
also before the training on their Day 1, to assess how good
the model is before seeing on the new force patterns. It is
worth stressing that Day 1data are used for (re)training, and
validations are always performed on Day 2data, assessing the
cross-day generalization capabilities of the trained model. The
described procedure is strictly incremental in that the model
never sees data from different sessions during a training stage:
we never perform multi-session or multi-day training.
2) Online Learning:In this work, we implement and evalu-
ate an online learning mode that mirrors the amount of data
and computational budget available in an embedded system
working in real time to process the stream of data produced by
a new sEMG acquisition session. For our real-time embedded
perspective, we define online learning as a (re)training of the
models that is performed with the following constraints:
•the training examples (i.e., 64 channels ×63 samples
windows) are consumed for training one-by-one: a full
forward-backward-update is computed exploiting a single
input window;
•this one-by-one procedure strictly happens in the order
of acquisition, feeding the input windows to the training
in the exact order they are recorded;
•this pass on the session is only done once, so that each
input window only contributes once to the training, i.e.,
with just one forward-backward-update pass (performed
individually for each window, due to the requirement of
the first point).
In our application, each example element is a 64 channels ×
63 samples sEMG time-window, taken with slide 64 samples,
i.e., 31.2 ms, as detailed in III-B. Therefore, in online learning,
each of these input windows is processed to perform one
iteration of gradient descent of the TCN, consisting of a
forward pass, a backward pass, and a model update. We
compare the online learning against a baseline represented by a
longer, computation-intensive offline training, more favorable
to accuracy but not feasible on-device due to the memory
footprint of complete training sets and the computational
burden of several epochs. In detail, we implement and compare
the following two training configurations:
•Offline baseline: randomized mini-batching, mini-batch
size 32, Adam optimizer with initial learning rate 1·10−4,
32 epochs (i.e., the whole training set is seen 32 times);
•Online: sequential mini-batching, mini-batch size 1, gra-
dient descent with fixed learning rate 2·10−4,1epoch
(i.e., just 1pass over the training set).
The online settings implement the constraints of real-time
training on a computational platform with limited memory:
each example is exploited for one forward-backward-update
as soon as it is acquired, then discarded; no second traversal
on the training session’s data is performed since an embedded
device is typically not able to store an sEMG session on board.
It is worth remarking that the incremental setup and the on-
line setup are orthogonal experimental settings. In particular,
the offline training baseline follows the incremental protocol
like the online training but does not implement the latter’s
streaming behavior. All (re)trainings have the Mean Squared
6 GENERIC COLORIZED JOURNAL, VOL. XX, NO. XX, XXXX 2017
FABRIC CONTROLLER (FC)’s
clock and voltage domain CLUSTER’s
clock and voltage domain
cluster
DMA
µDMA
FC
shared L1 memory
logarithmic interconnect
FPU
Master/Core 8
Core 0
Core 1
Core 2
Core 3
Core 4
Core 5
Core 6
Core 7
CLUSTER
GPIO
and
PWM RAM
64kB
1.5MB
interleaved
flash
2MB
interfaces
Fig. 3: High-level block diagram of the GAP9 MCU.
Error as a loss function. We always perform training on data
from HYSER’s Day 1and validation on Day 2(III-A), thus
obtaining a cross-day assessment that accounts for the sEMG’s
inherent cross-session variability mostly due to electrode repo-
sitioning [23], [24], [54]. We conduct all experiments within-
subject, without any multi-subject training or cross-subject
validation. In this way, we target the adaptation to new force
patterns and the cross-day generalization in a subject-specific
scenario. The sEMG’s inter-subject variability is outside the
scope of this work.
D. Deployment on a Parallel ULP MCU
We execute our application on the commercial microcon-
troller GWT GAP9.3It is equipped with a Parallel Ultra-
Low Power (PULP)4[56], [57] 9-core cluster accelerator
exploiting a RISC-V Instruction Set Architecture extended
with instructions specialized for digital signal processing and
ML (linear algebra). Fig. 3 shows a high-level block diagram
of the platform. GAP9 represents the SoA of low-power MCUs
since it has ranked first in latency and energy consumption in
the MLPerf Tiny v1.05benchmarking.
We have deployed the TCN model’s (III-B) learn-
ing via online back-propagation (III-C.2) by using the
TrainLib Deployer tool of the PULP-TrainLib6[58],
[59] library, the first open-source DNN training library for
RISC-V-based multi-core MCUs (such as GAP9, mounting
PULP), developed to enable the On-Device Learning on this
category of parallel ultra-low-power platforms. The library
offers matrix multiplication primitives for training and utilities
for back-propagation in float32 and float16 on multiple
cores. The library’s TrainLib Deployer is the automated
code generator for validating and training user-specified DNNs
on a PULP-based platform. The deployer’s output is a C source
code project to run and profile the training update stages of
the model, performing back-propagation with the mini-batch
size and a number of epochs set by the user.
3Product brief at https://greenwaves-technologies.com/
gap9_processor/; for a thorough exposition, refer to the paper about
GWT GAP8 [55], the predecessor of GAP9.
4https://pulp-platform.org/
5https://mlcommons.org/en/inference-tiny- 10/
6https://github.com/pulp-platform/pulp- trainlib
We performed the online training experiments by streaming
the sEMG windows of the HYSER trials from a PC to the
GAP9 MCU via a serial interface; GAP9 processes each
window executing one training step, i.e., forward pass, back-
ward pass, and weights update. GAP9 repeats the reception-
forward-backward-update loop online for each sEMG window
individually.
The profiling of the net’s training updates was conducted
by executing with GAP9’s most energy-efficient settings,
namely, Vdd core = 0.65 V and fCLK = 240 MHz. Latency was
measured in cycles exploiting the performance counter of the
API of PMSIS,7the open-source system layer for GAP9’s
operating system. Latency in physical time was determined
as num cycles/fCLK . The power consumption was measured
experimentally with GAP9’s Evaluation Kit8, 9 and a Nordic
Semiconductor Power Profiler Kit II (PPK2).10 We used the
PPK2 to measure the GAP9 cluster’s current draw (excluding
the peripherals and the off-chip memories), we used a GPIO to
synchronize the current measurement with the code execution,
and we determined the energy draw as power×latency.
IV. EXPERIMENTAL RESULTS
A. Regression Error
Fig. 4 shows the MAE obtained with the incremen-
tal learning protocol (III-C.1) on all trials of all 20 sub-
jects of the HYSER dataset, using the baseline offline
training and the online training (III-C.2). The MAEs of
Day 1of (1-DoF,Stage 0), of (5-DoF,Stage 1), and of
(1-RANDOM,Stage 2) are training MAEs, reported to check
the cross-day training-to-validation accuracy drop; all other
distributions are validation MAEs.
The most relevant difference between the offline (Fig. 4a)
and the online training (Fig. 4b) is that the latter yields
higher regression errors, both in training and validation. This
difference is expected since the online training works in
streaming by doing a single pass on the training set (i.e., only
1epoch) and has thus less opportunity to exploit the training
data. However, only the online setup is deployable on resource-
constrained embedded devices.
A second general trend, observed in both the offline and the
online training, is that in all validations (i.e., on Day 2), the
models retrained more recently (based on the protocol stages
illustrated in III-C.1) are more accurate. In detail:
•on Day 2of the 5-DoF dataset, the Stage-1TCN has a
lower error distribution than the Stage-0TCN since the
latter is specialized on the single-finger force patterns of
the 1-DoF dataset whereas the former has seen Day 1
of the 5-DoF dataset;
•analogously, on Day 2of the RANDOM dataset, the Stage-
2TCN has a lower error distribution than the Stage-1
7https://greenwaves-technologies.com/manuals/
BUILD/HOME/html/index.html
8https://greenwaves-technologies.com/product/gap9_
evk-gap9- evaluation-kit- efused/
9https://greenwaves-technologies.com/product/
gap9-resources/
10https://www.nordicsemi.com/Products/
Development-hardware/Power- Profiler-Kit- 2
M. ZANGHIERI et al.: PREPARATION OF PAPERS FOR IEEE TRANSACTIONS AND JOURNALS (FEBRUARY 2017) 7
day 1 day 2 day 1 day 2 day 1 day 2
1-DOF DATASET 5-DOF DATASET RANDOM DATASET
0.0
2.5
5.0
7.5
10.0
12.5
15.0
17.5
20.0
Mean absolute error [% MVC]
Stage 0
Stage 1
Stage 2
(a)
day 1 day 2 day 1 day 2 day 1 day 2
1-DOF DATASET 5-DOF DATASET RANDOM DATASET
0.0
2.5
5.0
7.5
10.0
12.5
15.0
17.5
20.0
Mean absolute error [% MVC]
Stage 0
Stage 1
Stage 2
(b)
Fig. 4: Experimental distributions of the MAE obtained for the incremental learning stages for baseline offline train-
ing and online training. 4a: Baseline offline training. 4b: Online training. The MAEs of (Day 1,1-DoF,Stage 0), of
(Day 1,5-DoF,Stage 1), and of (Day 1,RANDOM,Stage 2), are training MAEs, since Day 1is the training set of these
dataset-stage pairs, as explained in III-C.1; all other distributions are validation MAEs; the training metrics are reported to
check the cross-day training-to-validation accuracy drop. The lower (resp., upper) whisker is set at the lowest datum above
Q1−1.5·IQR (resp., Q3+ 1.5·IQR), with Q1and Q3the first and third quartiles respectively, and IQR ≜Q3−Q1the
interquartile range.
8 GENERIC COLORIZED JOURNAL, VOL. XX, NO. XX, XXXX 2017
TABLE IV: Results of the ablation experiments that reduce the
input time window, the HD-sEMG data rate, or the number
of HD-sEMG channels (one at a time). The baseline is the
setup using 63-sample input time-windows of 64 HD-sEMG
channels sampled at 2048 Hz (detailed in III-B; dedicated
results in IV-A).
Ablation experiment MAE, reported as
(median ±IQR) % MVC
input time window 47 samples 10.15 ±3.93
31 samples 10.54 ±4.02
data rate 1024 Hz 10.33 ±3.98
48 HD-sEMG channels 9.98 ±3.91
Reference setup (III-B; IV-A) 9.58 ±3.89
TCN (which in turn performs better than the Stage-0
TCN) since the Stage-2TCN has seen multi-fingers force
patterns out of the fixed combinations of 5-DoF during
retraining on RANDOM’s Day 1.
This result confirms the motivation for the computational
investment in retraining, showing that it is required to im-
prove the regression accuracy. It is worth observing that this
improvement is present for both the baseline offline training
and the online training, showing that the online training can
learn even with the online protocol’s limitations (4b).
The most significant validation for summarizing our
method’s accuracy and generalization ability is the cross-day
validation on HYSER RANDOM, which contains diverse non-
predefined force patterns like a real-world scenario. In this
cross-day validation, the baseline non-deployable offline train-
ing and the online training reach a median regression error of
(9.10 ±4.02)% MVC and (9.58 ±3.89)% MVC, respectively,
expressed as median ±Inter-Quartile Range (IQR). This result
indicates that the online training protocol produces a negligible
accuracy loss compared to the physiological variability range
across subjects and trials. Moreover, this regression error is
comparable to that attained by the SoA literature targeting
the HYSER dataset, which is in the range (8.5±5.0)%
MVC in simpler settings, such as within-day validation [29],
validation on 5-DoF [51], or validation on 1-DoF [53].
This comparison demonstrates that our setup matches the
SoA regression accuracy on the dataset. Moreover, the cited
regression works on HYSER only target the regression error
without addressing the deployment on embedded hardware and
the real-time operation.
B. Ablation Study
The previous subsection (IV-A) exposes the results of the
setup that uses 63-sample input time-windows of 64 HD-
sEMG channels sampled at 2048 Hz (detailed in III-B). We
identified these settings as a sweet spot providing an effective
trade-off between accuracy and a low resource budget. We
motivate this choice by reporting in Table IV the results
of ablation experiments that decrease the computation and
memory requirement by reducing the following parameters
(one at a time):
•input time window: reduced from 63 samples to 47 or 31
samples (amounting to 22.9 ms or 15.1 ms, respectively);
•data rate: subsampled from 2048 Hz to 1024 Hz;
•channels: reduced from 64 to 48; the 48 channels were
obtained by downsampling the 4HYSER’s 8×8sen-
sor patches by 2×and 3×longitudinally and trans-
versely with respect to the forearm, respectively (the in-
verse choice, i.e., 3×longitudinally and 2×transversely,
yielded worse accuracy).
These ablations were applied one at a time, i.e., one for
each ablation experiment. As can be seen from the results in
Table IV, all these more compact settings yield sub-optimal
regression errors that are not competitive with the reference
setup, motivating our choices. The following subsection ex-
poses the deployment and execution results that show that
the proposed reference setup is efficient and suitable for an
embedded implementation.
C. Profiling on a Parallel ULP MCU
The profiling of the memory occupation of each TCN’s
forward-backward-update step with mini-batch size 1yielded
110.9 KiB, which amounts to 86.6% of the 128 KiB available
in the L1 memory of GAP9’s cluster (III-D). This compact
memory footprint shows that our model combines high accu-
racy and a hardware-friendly size suitable for TinyML appli-
cations on resource-constrained devices. As to execution, the
profiling yielded a latency of 356 k cycles for each forward-
backward-update step of the whole net; this corresponds to
1.49 ms at fclk = 240 MHz, which is the clock frequency
of GAP9’s most energy-efficient configuration (III-D). This
latency is consistent with the training setting of consuming
one data window every 31.2 ms (III-B), indicating that our
application fulfills the requirements of real-time training. The
consensus for the latency requirement for sEMG-based HMIs
to make users perceive the control as real-time is 300 ms [60].
Taking into account our 30.8 ms input window plus the 1.5 ms
of computation latency, our application matches the sEMG
HMI real-time requirements, proving capable of providing
a fluid control without a significant perceived delay. The
measured power draw was 27.1 mW, which yields an energy
consumption of 40.4µJfor every forward-backward-update
step. This value shows that our solution fits the requirements
of low-power operation on embedded platforms.
V. DISCUSSION
This section discusses the contribution of this work in
comparison with the related SoA literature. In particular, our
contribution can be better identified based on the scheme we
provide in Table II, reporting the SoA works that use the
HYSER dataset targeting the regression task of HD-sEMG-
based force estimation. These works represent a minority since
most research on the HYSER dataset addresses its PR sub-
dataset for discrete recognition of fixed hand positions; this
limitation is common to the domain of sEMG-based hand
modeling in general [6].
On the HYSER dataset, the most relevant assessment to
determine a model’s accuracy and generalization robustness is
on the RANDOM dataset (featuring diverse forces not fixed by
protocol) in a cross-day validation to be close to a realistic
M. ZANGHIERI et al.: PREPARATION OF PAPERS FOR IEEE TRANSACTIONS AND JOURNALS (FEBRUARY 2017) 9
utilization. In these conditions, the proposed online training
achieves a regression error (9.58 ±3.89)% MVC (IV-A). This
regression error falls in the same range as the error achieved
in literature, which lies in the range of (8.5±5.0)% MVC
for validations conducted in settings that are easier from the
Machine Learning viewpoint, namely, validation performed
within-day [29], on 5-DoF [51], or on 1-DoF [53]. A
remarkable result is that the constraints introduced by the
online learning protocol only cause a negligible regression
error increase, compared to the physiological variability range
across subjects and trials (this physiological variability is an
established trend also observed in the literature, as can be seen
in the numerical results of Table II). Globally, the comparison
highlights that our proposed solution is competitive with the
SoA accuracy of sEMG-based force estimation.
As far as deployment and execution are concerned, we
are the first to report several progresses. The SoA literature
does not address the implementation on resource-constrained
hardware, which is a requirement of enabling wearable HMIs.
In contrast with this general limitation of all the cited
works, we present a TinyML solution that is deployed and
profiled an embedded computation device [36]–[38] suitable
for a low-power real-time wearable sEMG-based controller.
Moreover, our solution is more advanced compared to other
TinyML sEMG applications that, though successfully exploit-
ing resource-constrained embedded hardware, only deploy
inference [20], [24], [34]. In contrast, we also implement real-
time incremental online learning as a key contribution.
VI. CONCLUSION
In this work, we have presented an incremental online-
training solution for sEMG-based estimation of simultaneous
multi-finger forces. We propose a tiny TCN devised so that its
training computational budget fits the strict resource require-
ments of embedded platforms. We evaluate our strategy on the
HYSER dataset in a cross-day validation, accounting for the
sEMG’s inherent inter-session variability. With an incremental
training protocol, our online training reaches a cross-day MAE
of (9.58 ±3.89)% MVC on HYSER’s RANDOM subset, which
has no predefined force pattern sequences and is thus the
most challenging and closest to a real scenario. This error
is in the same range reached by an offline training with
more epochs and resources and is compatible with the SoA
literature tackling HYSER with easier training and validation
settings. We profile our TCN’s online training on the SoA
parallel ultra-low power microcontroller GAP9, obtaining a
latency of 1.49 ms and an energy consumption of 40.4µJ
for each forward-backward-update step. This proves that our
solution fits the requirements of real-time training-on-device
on an embedded platform. This work expands our previous
paper [25] by contributing a more advanced retraining strategy:
we progress from sparse low-count to high-count, high-density
electrodes, from classification of fixed gestures to regression,
and from cross-arm-posture adaptation to the learning of
new simultaneous multi-finger force patterns. This advance is
promising for the future intuitive and realiable non-invasive
wearable HMIs.
In future work, we will exploit this work’s insights about
the computational budget of incremental online learning for
HD-sEMG-based force regression to implement a high-count
sEMG setup and to realize a novel research dataset for hand
force modeling [29], [30].
ACKNOWLEDGMENT
We thank Davide Nadalini (University of Bologna, Poly-
technic of Turin) and Alberto Dequino (University of Bologna,
Polytechnic of Turin) for the assistance in using their PULP-
TrainLib library. We thank GreenWaves Technologies for the
preview access to the GAP9 SDK.
REFERENCES
[1] C. J. De Luca, “The use of surface electromyography in biomechanics,”
Journal of Applied Biomechanics, vol. 13, no. 2, pp. 135–163, May
1997. DOI:10.1123/jab.13.2.135.
[2] R. M. Rangayyan, Biomedical Signal Analysis. Wiley, Apr. 2015. DOI:
10.1002/9781119068129.
[3] S. Benatti et al., “Multiple biopotentials acquisition system for
wearable applications,” in Proceedings of the International Joint
Conference on Biomedical Engineering Systems and Technologies -
Volume 1, ser. BIOSTEC 2015, SCITEPRESS - Science and Tech-
nology Publications, Lda, 2015, pp. 260–268. DO I:10 . 5220 /
0005320302600268.
[4] C. Castellini and P. van der Smagt, “Surface EMG in advanced hand
prosthetics,” Biological Cybernetics, vol. 100, no. 1, pp. 35–47, Nov.
2008. DOI:10.1007/s00422-008-0278-1.
[5] L. Guo et al., “Human-machine interaction sensing technology based
on hand gesture recognition: A review,” IEEE Transactions on Human-
Machine Systems, vol. 51, no. 4, pp. 300–309, 2021. DOI:10.1109/
THMS.2021.3086003.
[6] A. Phinyomark and E. Scheme, “EMG pattern recognition in the era
of big data and deep learning,” Big Data and Cognitive Computing,
vol. 2, no. 3, 2018. DOI:10.3390/bdcc2030021.
[7] Y. Wei et al., “A review of algorithm & hardware design for AI-based
biomedical applications,” IEEE Transactions on Biomedical Circuits
and Systems, vol. 14, no. 2, pp. 145–163, 2020. DOI:10 . 1109 /
TBCAS.2020.2974154.
[8] R. Meattini et al., “Experimental evaluation of a sEMG-based human-
robot interface for human-like grasping tasks,” in 2015 IEEE Inter-
national Conference on Robotics and Biomimetics (ROBIO), 2015,
pp. 1030–1035. DO I:10.1109/ROBIO.2015.7418907.
[9] S. Benatti et al., “Analysis of robust implementation of an EMG
pattern recognition based control,” in Proceedings of the International
Joint Conference on Biomedical Engineering Systems and Technolo-
gies - Volume 4, ser. BIOSTEC 2014, SCITEPRESS - Science and
Technology Publications, Lda, 2014, pp. 45–54. DOI:10 . 5220 /
0004800300450054.
[10] B. Milosevic et al., “Exploring arm posture and temporal variability in
myoelectric hand gesture recognition,” in 2018 7th IEEE International
Conference on Biomedical Robotics and Biomechatronics (BioRob),
2018, pp. 1032–1037. DOI:10.1109/BIOROB.2018.8487838.
[11] V. H. Cene et al., “Open database for accurate upper-limb intent detec-
tion using electromyography and reliable extreme learning machines,”
Sensors, vol. 19, no. 8, 2019. DOI :10.3390/s19081864.
[12] S. Benatti et al., “Online learning and classification of EMG-based
gestures on a parallel ultra-low power platform using hyperdimensional
computing,” IEEE Transactions on Biomedical Circuits and Systems,
vol. 13, no. 3, pp. 516–528, 2019. DO I:10.1109 /TBCAS. 2019.
2914476.
[13] E. Donati et al., “Discrimination of EMG signals using a neuromorphic
implementation of a spiking neural network,” IEEE Transactions on
Biomedical Circuits and Systems, vol. 13, no. 5, pp. 795–803, 2019.
DOI:10.1109/TBCAS.2019.2925454.
[14] A. Vitale et al., “Neuromorphic edge computing for biomedical ap-
plications: Gesture classification using EMG signals,” IEEE Sensors
Journal, vol. 22, no. 20, pp. 19 490–19 499, 2022. D OI:10 .1109 /
JSEN.2022.3194678.
[15] A. Krasoulis and K. Nazarpour, “Myoelectric digit action decoding
with multi-output, multi-class classification: An offline analysis,” Sci-
entific Reports, vol. 10, no. 1, Oct. 2020. DO I:10.1038/s41598-
020-72574- 7.
10 GENERIC COLORIZED JOURNAL, VOL. XX, NO. XX, XXXX 2017
[16] Y. Hu et al., “A novel attention-based hybrid CNN-RNN architecture
for sEMG-based gesture recognition,” PLOS ONE, vol. 13, pp. 1–18,
Oct. 2018. DO I:10.1371/journal.pone.0206049.
[17] E. Ceolini et al., “Hand-gesture recognition based on EMG and
event-based camera sensor fusion: A benchmark in neuromorphic
computing,” Frontiers in Neuroscience, vol. 14, 2020. DOI :10.3389/
fnins.2020.00637.
[18] S. Tam et al., “A fully embedded adaptive real-time hand gesture
classifier leveraging HD-sEMG and deep learning,” IEEE Transactions
on Biomedical Circuits and Systems, vol. 14, no. 2, pp. 232–243, 2020.
DOI:10.1109/TBCAS.2019.2955641.
[19] F. Chamberland et al., “Novel wearable HD-EMG sensor with shift-
robust gesture recognition using deep learning,” IEEE Transactions on
Biomedical Circuits and Systems, vol. 17, no. 5, pp. 968–984, 2023.
DOI:10.1109/TBCAS.2023.3314053.
[20] N. Leroux et al., “Online transformers with spiking neurons for
fast prosthetic hand control,” in 2023 IEEE Biomedical Circuits and
Systems Conference (BioCAS), 2023, pp. 1–6. DOI:10 . 1109 /
BioCAS58349.2023.10388996.
[21] E. Donati et al., “Long-term stable electromyography classification
using canonical correlation analysis,” in 2023 11th International
IEEE/EMBS Conference on Neural Engineering (NER), 2023, pp. 1–4.
DOI:10.1109/NER52421.2023.10123768.
[22] P. Kaufmann et al., “Fluctuating EMG signals: Investigating long-term
effects of pattern matching algorithms,” in 2010 Annual International
Conference of the IEEE Engineering in Medicine and Biology, 2010,
pp. 6357–6360. DO I:10.1109/IEMBS.2010.5627288.
[23] S. Ams ¨
uss et al., “Long term stability of surface EMG pattern
classification for prosthetic control,” in 2013 35th Annual International
Conference of the IEEE Engineering in Medicine and Biology Society
(EMBC), IEEE, 2013. DO I:10.1109/embc.2013.6610327.
[24] M. Zanghieri et al., “Robust real-time embedded EMG recognition
framework using temporal convolutional networks on a multicore IoT
processor,” IEEE Transactions on Biomedical Circuits and Systems,
vol. 14, no. 2, pp. 244–256, 2020. DO I:10.1109 /TBCAS. 2019.
2959160.
[25] M. Zanghieri et al., “Online unsupervised arm posture adaptation for
sEMG-based gesture recognition on a parallel ultra-low-power micro-
controller,” in 2023 IEEE Biomedical Circuits and Systems Conference
(BioCAS), 2023, pp. 1–5. DO I:10. 1109/ BioCAS58349. 2023.
10388902.
[26] M. Zanghieri et al., “Temporal variability analysis in sEMG hand grasp
recognition using temporal convolutional networks,” in 2020 2nd IEEE
International Conference on Artificial Intelligence Circuits and Systems
(AICAS), 2020, pp. 228–232. DO I:10.1109/AICAS48895.2020.
9073888.
[27] A. Burrello et al., “Tackling time-variability in sEMG-based gesture
recognition with on-device incremental learning and temporal convolu-
tional networks,” in 2021 IEEE Sensors Applications Symposium (SAS),
2021, pp. 1–6. DO I:10.1109/SAS51076.2021.9530007.
[28] A. Krasoulis et al., “Effect of user practice on prosthetic finger control
with an intuitive myoelectric decoder,” Frontiers in Neuroscience,
vol. 13, 2019. DOI:10.3389/fnins.2019.00891.
[29] X. Jiang et al., “Open access dataset, toolbox and benchmark pro-
cessing results of high-density surface electromyogram recordings,”
IEEE Transactions on Neural Systems and Rehabilitation Engineering,
vol. 29, pp. 1035–1046, 2021. DO I:10 . 1109 / TNSRE . 2021 .
3082551.
[30] A. Matran-Fernandez et al., “SEEDS, simultaneous recordings of high-
density emg and finger joint angles during multiple hand movements,”
Scientific Data, vol. 6, no. 1, Sep. 2019. DOI:10.1038/s41597-
019-0200- 9.
[31] P. Koch et al., “Regression of hand movements from sEMG data
with recurrent neural networks,” in 2020 42nd Annual International
Conference of the IEEE Engineering in Medicine & Biology Society
(EMBC), 2020, pp. 3783–3787. DO I:10 . 1109 / EMBC44109 .
2020.9176278.
[32] T. Bao et al., “A deep Kalman filter network for hand kinematics
estimation using sEMG,” Pattern Recognition Letters, vol. 143, pp. 88–
94, 2021. DO I:10.1016/j.patrec.2021.01.001.
[33] L. Meng et al., “Evaluation of decomposition parameters for high-
density surface electromyogram using fast independent component
analysis algorithm,” Biomedical Signal Processing and Control,
vol. 75, p. 103 615, 2022. D OI:10.1016/j.bspc.2022.103615.
[34] M. Zanghieri et al., “sEMG-based regression of hand kinematics
with temporal convolutional networks on a low-power edge micro-
controller,” in 2021 IEEE International Conference on Omni-Layer
Intelligent Systems (COINS), 2021, pp. 1–6. DOI:10 . 1109 /
COINS51742.2021.9524188.
[35] M. Zanghieri et al., “Event-based low-power and low-latency re-
gression method for hand kinematics from surface EMG,” in 2023
9th International Workshop on Advances in Sensors and Interfaces
(IWASI), 2023, pp. 293–298. DOI:10.1109/IWASI58316.2023.
10164372.
[36] L. Dutta and S. Bharali, “TinyML meets IoT: A comprehensive
survey,” Internet of Things, vol. 16, p. 100461, 2021. DOI:10.1016/
j.iot.2021.100461.
[37] P. P. Ray, “A review on TinyML: State-of-the-art and prospects,”
Journal of King Saud University - Computer and Information Sciences,
vol. 34, no. 4, pp. 1595–1623, 2022. DOI:10.1016 /j .jksuci .
2021.11.019.
[38] Y. Abadade et al., “A comprehensive survey on TinyML,” IEEE
Access, pp. 1–1, 2023. DOI:10.1109/ACCESS.2023.3294111.
[39] L. Ravaglia et al., “Memory-latency-accuracy trade-offs for continual
learning on a RISC-V extreme-edge node,” in 2020 IEEE Workshop
on Signal Processing Systems (SiPS), 2020, pp. 1–6. DOI:10.1109/
SiPS50750.2020.9195220.
[40] C. Lea et al., “Temporal convolutional networks for action segmenta-
tion and detection,” in 2017 IEEE Conference on Computer Vision and
Pattern Recognition (CVPR), 2017, pp. 1003–1012. DOI:10.1109/
CVPR.2017.113.
[41] S. Bai et al., “An empirical evaluation of generic convolutional and re-
current networks for sequence modeling,” CoRR, vol. abs/1803.01271,
2018. DOI:10.48550/arXiv.1803.01271.
[42] A. Burrello et al., “TCN mapping optimization for ultra-low power
time-series edge inference,” in 2021 IEEE/ACM International Sympo-
sium on Low Power Electronics and Design (ISLPED), 2021, pp. 1–6.
DO I:10.1109/ISLPED52811.2021.9502494.
[43] M. Zanghieri et al., “Low-latency detection of epileptic seizures from
iEEG with temporal convolutional networks on a low-power parallel
MCU,” in 2021 IEEE Sensors Applications Symposium (SAS), 2021,
pp. 1–6. DO I:10.1109/SAS51076.2021.9530181.
[44] A. Burrello et al., “Embedding temporal convolutional networks for
energy-efficient PPG-based heart rate monitoring,” ACM Trans. Com-
put. Healthcare, vol. 3, no. 2, Mar. 2022. DO I:10.1145/3487910.
[45] Y. Li et al.,Revisiting batch normalization for practical domain
adaptation, 2016. DO I:10.48550/arXiv.1603.04779.
[46] Y. Du et al., “Surface EMG-based inter-session gesture recognition
enhanced by deep domain adaptation,” Sensors, vol. 17, no. 3, 2017.
DO I:10.3390/s17030458.
[47] L. Pellegrini et al., “Latent replay for real-time continual learning,”
in 2020 IEEE/RSJ International Conference on Intelligent Robots
and Systems (IROS), 2020, pp. 10 203–10 209. DOI:10 . 1109 /
IROS45743.2020.9341460.
[48] C. Castellini et al., “Fine detection of grasp force and posture by
amputees via surface electromyography,” Journal of Physiology -
Paris, vol. 103, no. 3, pp. 255–262, 2009. DO I:10 . 1016 / j .
jphysparis.2009.08.008.
[49] R. C. Sˆ
ımpetru et al., “Accurate continuous prediction of 14 degrees
of freedom of the hand from myoelectrical signals through convolutive
deep learning,” in 2022 44th Annual International Conference of the
IEEE Engineering in Medicine & Biology Society (EMBC), 2022,
pp. 702–706. DO I:10.1109/EMBC48229.2022.9870937.
[50] R. C. Sˆ
ımpetru et al., “Sensing the full dynamics of the human hand
with a neural interface and deep learning,” bioRxiv, 2022. DOI:10 .
1101/2022.07.29.502064.
[51] X. Jiang et al., “Random channel masks for regularization of least
squares-based finger EMG-force modeling to improve cross-day per-
formance,” IEEE Transactions on Neural Systems and Rehabilitation
Engineering, vol. 30, pp. 2157–2167, 2022. DO I:10.1109/TNSRE.
2022.3194246.
[52] X. Jiang et al., “Explainable and robust deep forests for EMG-
force modeling,” IEEE Journal of Biomedical and Health Informatics,
vol. 27, no. 6, pp. 2841–2852, 2023. DO I:10.1109/JBHI.2023.
3262316.
[53] W. Wu et al., “A new EMG decomposition framework for upper limb
prosthetic systems,” Journal of Bionic Engineering, Jul. 2023. DOI:
10.1007/s42235-023- 00407-0.
[54] F. Palermo et al., “Repeatability of grasp recognition for robotic
hand prosthesis control based on sEMG data,” in 2017 International
Conference on Rehabilitation Robotics (ICORR), 2017, pp. 1154–1159.
DO I:10.1109/ICORR.2017.8009405.
[55] E. Flamand et al., “GAP-8: A RISC-V SoC for AI at the edge of
the IoT,” in 2018 IEEE 29th International Conference on Application-
M. ZANGHIERI et al.: PREPARATION OF PAPERS FOR IEEE TRANSACTIONS AND JOURNALS (FEBRUARY 2017) 11
specific Systems, Architectures and Processors (ASAP), 2018, pp. 1–4.
DOI:10.1109/ASAP.2018.8445101.
[56] F. Conti et al., “PULP: A ultra-low power parallel accelerator for
energy-efficient and flexible embedded vision,” J. Signal Process. Syst.,
vol. 84, no. 3, pp. 339–354, Sep. 2016. DOI:10.1007/s11265-
015-1070- 9.
[57] A. Garofalo et al., “PULP-NN: Accelerating quantized neural net-
works on parallel ultra-low-power RISC-V processors,” Philosophical
Transactions of the Royal Society A: Mathematical, Physical and
Engineering Sciences, vol. 378, no. 2164, p. 20 190 155, Dec. 2019.
DOI:10.1098/rsta.2019.0155.
[58] D. Nadalini et al., “PULP-TrainLib: Enabling on-device training for
RISC-V multi-core MCUs through performance-driven autotuning,” in
Embedded Computer Systems: Architectures, Modeling, and Simula-
tion, Cham: Springer International Publishing, 2022, pp. 200–216. DO I:
10.1007/978-3- 031-15074- 6_13.
[59] D. Nadalini et al., “Reduced precision floating-point optimization for
deep neural network on-device learning on microcontrollers,” Future
Generation Computer Systems, vol. 149, pp. 212–226, 2023. DOI :
https://doi.org/10.1016/j.future.2023.07.020.
[60] B. Hudgins et al., “A new strategy for multifunction myoelectric
control,” IEEE Transactions on Biomedical Engineering, vol. 40, no. 1,
pp. 82–94, 1993. DOI:10.1109/10.204774.
Marcello Zanghieri (Graduate Student Mem-
ber, IEEE) received his M.Sc. in Physics (cum
laude) from the University of Bologna, Italy, in
2019. He is currently working toward his Ph.D.
in Data Science and Computation under the
supervision of Prof. Luca Benini at the Energy-
Efficient Embedded Systems Laboratory (EEES
Lab) of the University of Bologna. His research
interests focus on time series analysis with ma-
chine learning and deep learning, focusing on
sEMG, ultrasounds, and EEG to advance real-
time human-machine interaction based on embedded computing plat-
forms, including parallel (ultra-)low-power microcontrollers.
Pierangelo Maria Rapa (Graduate Student
Member, IEEE) received his M.Sc. degree in
Electronic Engineering from the University of
Bologna, Italy, in 2022. He is currently work-
ing on his Ph.D. in Automotive Engineering
for Intelligent Mobility under the supervision of
Prof. S. Benatti at the DEI department, Univer-
sity of Bologna. His research interests include
biosignals and ultra-low-power Human-Machine
Interfaces with specific emphasis on their appli-
cation in the automotive industry.
Mattia Orlandi (Graduate Student Member,
IEEE) received his M.Sc. degree in Artificial
Intelligence from the University of Bologna, Italy,
in 2022. He is currently working toward his
Ph.D. Data Science and Computation under the
supervision of Prof. S. Benatti at the Energy-
Efficient Embedded Systems Laboratory (EEES
Lab), DEI Department, University of Bologna.
His research activities involve bio-signal pro-
cessing with machine learning on low-power
computing platforms. He is investigating how to
decode EMG signals into spike trains to develop advanced human-
machine interfaces.
556 IEEE JOURNAL ON EMERGING AND SELECTED TOPICS IN CIRCUITS AND SYSTEMS, VOL. 10, NO. 4, DECEMBER 2020
Giacomo Indiveri (Senior Member, IEEE) received
the M.Sc. degree in electrical engineering and the
Ph.D. degree in computer science and electrical
engineering from the University of Genova, Genova,
Italy, in 1992 and 2004, respectively.
He was a Post-Doctoral Research Fellow with the
Division of Biology, Caltech, and with the Institute
of Neuroinformatics, University of Zürich and ETH
Zürich. He is currently a Dual Professor with the
Faculty of Science, University of Zürich, and with
the Department of Information Technology and Elec-
trical Engineering, ETH Zürich, Switzerland. He is also the Director of the
Institute of Neuroinformatics (INI), University of Zürich and ETH Zürich. His
research interests include the study of real and electronic neural processing
systems, with a particular focus on spike-based learning and spike-based
recurrent neural network dynamics. His research and development activities
focus on the full custom hardware implementation of real-time sensory-motor
systems using analog/digital neuromorphic circuits and emerging memory
technologies. He received the ERC Starting Grant on “Neuromorphic Proces-
sors” in 2011, and the ERC Consolidator Grant on neuromorphic cognitive
agents in 2016. He is a member of several technical committees of the IEEE
Circuits and Systems Society and a Fellow of the European Research Council.
Elisa Donati (Member, IEEE) received the B.Sc.
and M.Sc. degrees (cum laude) in biomedical engi-
neering from the University of Pisa, Pisa, Italy, and
the Ph.D. degree in biorobotics from the Sant’Anna
School of Advanced Studies, Pisa.
She is currently a Research Fellow with the
Institute of Neuroinformatics, University of Zürich
and ETH Zürich. Her research activities include
the interface of the neuroscience and neuromorphic
engineering. She is interested in understanding how
the biological neural circuits carry out the compu-
tation and apply them in biomedical application and neurorobotics. She is
investigating how to process biomedical signal to extract features to develop
human–robot machine. She is the Co-Coordinator of the H2020 EU CSA
Project NEUROTECH.
Authorized licensed use limited to: Universita degli Studi di Bologna. Downloaded on September 19,2023 at 12:05:17 UTC from IEEE Xplore. Restrictions apply.
Elisa Donati (Member, IEEE) received the
B.Sc. and M.Sc. degrees (cum laude) in biomed-
ical engineering from the University of Pisa,
Pisa, Italy, and the Ph.D. degree in biorobotics
from the Sant’Anna School of Advanced Studies,
Pisa. She is currently a Research Fellow with
the Institute of Neuroinformatics, University of
Z¨
urich and ETH Z ¨
urich. Her research activities
include the interface of the neuroscience and
neuromorphic engineering. She is interested in
understanding how the biological neural circuits
carry out the computations and apply them in biomedical application
and neurorobotics. She was the Co-Coordinator of the H2020 EU CSA
Project NEUROTECH.
Luca Benini (Fellow, IEEE) holds the chair
of digital Circuits and systems at ETH Zurich,
Switzerland, and is Full Professor at the Uni-
versit`
a di Bologna, Italy. He received his Ph.D.
from Stanford University. Dr. Benini’s research
interests are energy-efficient parallel computing
systems, smart sensing micro-systems, and ma-
chine learning hardware. He is a Fellow of the
ACM and a member of the Academia Europaea.
He is the recipient of the 2016 IEEE CAS Mac
Van Valkenburg Award, the 2020 EDAA Achieve-
ment Award, the 2020 ACM/IEEE A. Richard Newton Award, and the
2023 IEEE CS E.J. McCluskey Award.
Simone Benatti (Member, IEEE) received his
Ph.D. degree in Electronics, Telecommunica-
tions, and Information Technologies from the
University of Bologna under the supervision of
Prof. Luca Benini. During the Ph.D., he was a
visiting fellow at BWRC - University of California,
Berkeley (supervisor Prof. Jan Rabaey). Cur-
rently, he serves as Assistant Professor at the
University of Modena and Reggio Emilia while
pursuing his collaboration with the EEES Lab
of the University of Bologna. In 2023, he was
appointed Visiting Professor at EFCL-ETHZ. His research interests
focus on energy-efficient embedded systems for IoT and biomedical
applications. This includes hardware/software codesign to efficiently
address performance, as well as advanced algorithms. Dr. Benatti
works on designing and optimizing energy-efficient embedded systems
for biopotential (EMG and EEG) acquisition and processing and on
Brain-Machine Interfaces for Human-Machine Interaction. In this field,
he has published more than 90 papers in international peer-reviewed
conferences and journals. He has ongoing collaborations with several
international research institutes, such as ETHZ-EFCL, EPFL, TU Graz,
FBK, and Politecnico di Torino. Dr. Benatti is recipient of the GHAIA grant
(H2020-MSCA-RISE-2017, g.a. 777822).
