| KEYWORD |
Self-Supervised Incremental Learning for Microsensors in the Field
keywords INCREMENTAL LEARNING, MACHINE LEARNING, ARTIFICIAL NEURAL NETWORKS, MICRO NEURAL PROCESSOR, ONLINE LEARNING, SELF-SUPERVISED LEARNING, SENSOR DRIFT
Reference persons LUCIANO LAVAGNO, MIHAI TEODOR LAZARESCU
External reference persons Ing. Roberto Simmarano dell'azienda Sensichips s.r.l.
Research Groups Microelectronics, Sensing and processing
Thesis type COMPANY STAGE
Description Microsensors can measure the composition of gases and liquids, among other physical phenomena, and are widely used in various applications, such as environmental monitoring, industrial automation, healthcare, and security. They can provide precise and reliable data, making them invaluable for various industries, but training machine learning models solely in controlled laboratory environments may not consider the diversity and complexity of real-world field conditions, which results in suboptimal performance and decreased accuracy. To overcome this, microsensors should adapt and enhance their data processing autonomously in the field, without human intervention or supervision.
The aim of this thesis is to design and test unsupervised incremental learning techniques for microsensors, allowing them to learn self-supervised from data streams in the field while avoiding catastrophic forgetting (i.e., without losing learned information upon learning new one, which can severely impair its performance and functionality). Normally, no external intervention or supervision, human or machine, is possible in the field.
The proposed techniques will be based on artificial neural networks, with a regard to compatibility with the proprietary micro neural processor developed by Sensichips s.r.l., using MATLAB or Python frameworks, then possibly integrated with the Weka environment compatible with the company’s SLM-Studio suite for verification with field data. The performance of the algorithms will be evaluated using classification models, training and test data provided by the company, which include sensor drift scenarios. The expected outcome is to achieve improved accuracy, robustness, and adaptability of the microsensors in the field, compared to the initial models trained in the laboratory.
See also http://www.sensichips.com/
Required skills No specific skills are required.
Useful skills include sensor technology (principles and applications, causes and effects of sensor drift), data analysis (statistical techniques, evaluate accuracy and precision), machine learning basics, and microelectronics (microprocessor architectures).
Notes The thesis will be carried out in collaboration with Sensichips s.r.l., specialized in the development of highly miniaturized learning microsensors. Its sensors are based on an innovative microelectronic multi-sensor chip platform code-named SENIPLUS.
Deadline 24/11/2024
PROPONI LA TUA CANDIDATURA