Research Projects
Welcome to my research projects section. Here you can find an overview of my latest scientific contributions and interests. I specialize in neuromorphic technologies and their applications. Click on the project name to learn more.
Embedded Event-Based Computing for Hand Kinematics Regression
The sEMG-based regression is still scarcely explored in research since most approaches have addressed classification. Combining event-based EMG encoding and low-power microcontroller (STM32 F401, mounting ARM Cortex-M4) allows for fast and efficient regression of hand kinematics. Event-based encoding exploits neuromorphic computing and brings benefit in terms of latency and power consumption. Preliminary promising results shows
regression with Mean Absolute Error of 8.8 ± 2.3 degrees on 5 degrees of actuation on the public dataset NinaPro DB8, comparable with the SoA Deep Neural Network (DNN). At the same time, results shows 9× less memory and 13× less energy per inference, with 10× shorter latency per inference compared to the SoA deep net, proving suitable for resource-constrained embedded platforms.
Motorneurons analysis and classification
Muscles receive a neural activation signal from the pool of innervating motor neurons as the sum of the spiking activities of single motor neurons. Accessing motoneurons activity enables a more natural interaction allowing control of the force of muscular contractions. This approach provides information on the contraction force and not only its discrete approximation so that a continuous, more natural control is able to be generated. Blind Source Separation (BSS) algorithms allow the motoneurons decomposition from HD-sEMG. Among BSS, Fast-ICA is one common method to extract this info. However, its implementation is quite heavy and its embeddability is still an open challenge. Despite, the decomposition methods, once the motorneurons are generated, a natural way of processing the resulting spike trains is by using spiking neural networks that allow the processing of the info directly in the spike domain. As the first step, spike trains can be processed using a spiking convolutional neural network, as a combination of standard deep learning approach and spiking domain, showing promising results for future, more dedicated, implementation.
Hand Gesture classification using electromyographic signals
Processing superficial Electromyographic (EMG) signals to classify hand gestures using neuromorphic computing with the final goal of control prosthetic devices. The hardware spiking neural network (SNN), implemented on a general-purpose neuromorphic chip should recognise motion intention to use as input to generate commands for a prosthetic device.
This approach is completely different from standard machine learning as it removes the need for data transmission to external computing platforms. Different neuromorphic chips have been exploited during my research, both mixed-signal and digital showing great results in terms of accuracy, power consumption, and latency.
Different network topologies have been implemented (feed-forward networks, spiking neural network, delay chain ) and investigated based on the task and signal under investigation. The last layer, the readout layer, of each network was trained using classic methods (e.g. linear regression, support vector machine) or using spiking local learning rules (e.g. delta rule, Spike-timing-dependent plasticity). The results are always compared to traditional machine learning methods.
Event-Based Sensor Fusion
Multi-sensor data fusion mechanisms have been investigated to improve discrimination accuracy. In human-machine interfaces signals as electromyography (EMG) signal visual information can be integrated for improving classification accuracy. In traditional machine-learning, this multi-sensor approach, while improving accuracy and robustness, introduces the disadvantage of high computational cost, which grows exponentially with the number of sensors and the number of measurements. A fully neuromorphic sensor fusion approach can solve these limitations as showed by the integration of a event-based vision sensor and three different neuromorphic processor, i.e. Loihi and ODIN+MorphIC. The EMG signals were recorded using traditional electrodes and then converted into spikes to be fed into the chips. The fully neuromorphic approach was compared to a baseline implemented using traditional machine learning approaches on a portable GPU system.
According to the chips constraints, specific spiking neural networks (SNNs) for sensor fusion were designed and showed classification accuracy comparable to the software baseline. These neuromorphic alternatives have increased inference time, between 20% and 40%, with respect to the GPU system but have a significantly smaller energy-delay product (EDP) which makes them between 30x and 600x more efficient.
Long-term stable Electromyography classification
Long-term real-world application scenarios is challenging due to EMG variability (i.e. electrodes shift, muscle artifacts, fatigue, user adaptation, or skin-electrode interfacing issues). One of the most critical challenges is maintaining high EMG data classification performance across multiple days without retraining the decoding system. A novel statistical method based on canonical correlation analysis (CCA) showed excellent results in stabilizing EMG classification performance across multiple days. CCA can dramatically decrease the performance drop of standard classifiers observed across days, by maximizing the correlation among multiple-day acquisition data sets. Results show how the performance of a classifier trained on EMG data acquired only of the first day of the experiment maintains 90% relative accuracy across multiple days, compensating for the EMG data variability. This approach eliminates the need for large data sets and multiple or periodic training sessions, which currently hamper the usability of conventional pattern recognition based approaches.
Neuromorphic Twin for somatosensory cortex emulation
In bidirectional prosthetic device, recent studies showed that invasive neurostimulation produced a very fast adaptation (or habituation) phenomenon, which caused the stop of perceived sensations. To understand the phenomena behind the aforementioned adaptation and the involved neural mechanism a neuromorphic twin can be used to understand the ascending neural pathways and define the location and the onset of the adaptation.
By investigating the computational principles of cortical neural networks we will determine the mechanisms that can reproduce the observed adaptation phenomena and match the experimental data. The network so developed will include information from the peripheral nervous system, such as the spinal cord synapses that modulate the input to the network.
Event-Based Motor Control
The design of robots that interact autonomously with the environment and exhibit complex behaviors is an open challenge that can benefit from understanding what makes living beings fit to act in the world. One of the key worlds in motor control is closed-loop, where the agent to efficiently move in the environment needs to perceive it and take action based on it and its state. A closed loop happens at different levels, at a low level with the correct activation of a single motor, or at a high level where multiple joints are coordinated to allow the agent to move. By combining computational primitives it is possible to implement a motor controller on a mixed-signal analog/digital neuromorphic processor that continuously calculates an error signal from the desired target and the feedback signals and performs a proportional derivative controller (PID) and to solve the inverse kinematics of a multi-joint arm.
Silicon neurons to emulate dynamics of biological neurons
Design of a mixed-signal sub-threshold VLSI chip with 32 conductance-based neurons (Hodgkin – Huxley) and 32x16 non-plastic synapses. Each neuron can be used to faithfully reproduce the behavior of biological neurons, or to reproduce oscillatory behavior, as the central pattern generator using measured physiological data as the input. The device integrates physiological feedback from arterial O2, lung inflation, and blood pressure sensors and it can provide breath-to-breath modulation to the cardiac device activation. The neuron showed incredible results in reproducing the same membrane voltage trace of a biological neuron located in the brain stem. This result has received large media exposure with 477 pieces of coverage across radio, online, and print. UK outlets such as BBC News, Financial Times, The Times, The Guardian, Telegraph, Mail Online, Independent, and Express have reported the story and many other national papers worldwide and specialist science outlets including The Engineer, IFL Science & MIT Technology Review.