Davide Brunelli

Neural Architecture Search (NAS) methods have been growing in popularity. These techniques have been fundamental to automate and speed up the time-consuming and error-prone process of synthesizing novel Deep Learning (DL) architectures. In general, the synthesized Neural Networks architectures are too complex to be deployed in resource-limited IoT platforms since the architecture search is targeted to increase network accuracy a performance. This thesis project will study an innovative NAS search strategy targeted to trade-off execution latency, energy consumption, and memory footprint.

For the past two decades, many successful microbial fuel cell (METs) applications, such as bioenergy generation, environmental monitoring, resource recovery, and platform chemicals production, have been demonstrated. A promising technology is represented by the so call Plant-Microbial Fuel Cells (MFCs), a particular type of bioreactor that exploits symbiotic plants to increase the energy generated. 

The thesis project will investigate the feasibility, creation, and characterization of innovative inkjet-printed electrodes to create environmentally friendly MFCs exploiting the state-of-the-art Ceradrop F-Serie deposition platform.

While the rapid growth of the Internet of Things has brought great benefits to people's everyday life, the accompanying generation of large-scale electronic waste is becoming a global concern. Eco-friendly and disposable electronics, which at the end of the life cycle can the trashed as a normal package without impact on the environment, are envisioned as the most effective solution to the growing issue of electronic waste. 

This thesis project will encompass the study and analysis of printed flexible electronics and devices exploiting the DII state-of-the-art Ceradrop F-Serie hybrid materials deposition platform combining Inkjet and Aerosol Jet technologies.

Le reti neurali possono essere utilizzate in sistemi di controllo qualità di processi industriali, ad esempio per individuare difettosità con un controllo visivo automatico. I recenti sviluppi tecnologici consentono di ottimizzare l’esecuzione di una rete neurale in modo che l’inferenza possa essere eseguita direttamente nella camera che acquisisce l’immagine.

Lo stage proposto consiste nell’ottimizzare la classificazione dei difetti di un processo di stampaggio plastico industriale tramite controllo visivo con camere programmabili: le reti neurali vengono eseguite in un microprocessore STM32 e comunicano il risultato della classificazione ad un gateway industriale che controlla anche l’avanzamento di un nastro trasportatore.

Il sistema è controllato remotamente e comunica con un cloud Azure.

Lo stage verrà svolto presso RosaMicro spa, azienda leader nella componentistica plastica e in collaborazione con STMicroelectronics, che fornisce sensori, microcontrollori e sistemi HW/SW per l’elaborazione dei segnali in ambito Internet of Things (IoT).

Static deep learning models are exposed to context drift, thus implying a model re-training. This procedure can be inefficient, especially in production chain processes. Introducing on-line learning capabilities to traditional deep learning models overcomes the context drift problem. 

This thesis aims to implement and test novel on-line learning solutions with unsupervised data and deploy them on edge devices.

Deep learning models can solve complex tasks such as detection and recognition. However, deploying such models in resource-constrained edge devices is challenging because of their complex architecture, thus, a high number of parameters. 

Knowledge distillation is a novel paradigm that aims to transfer knowledge from a teacher network to a student network that presents fewer parameters. With this approach, it is possible to deploy complex deep learning models in edge devices with a negligible drop in accuracy. 

This thesis aims to investigate and implement knowledge distillation solutions, test them with state-of-the-art datasets such as ImageNet or Visual Wake Words, and deploy the resulting student model in edge devices.

Quantization is a fundamental step to optimize deep neural networks to fit memory and processing constraints. The most used technique is post-training quantization; however, it can considerably decrease accuracy. 

In-training quantization is a novel technique that quantizes deep models during the training procedure and can lead to better performance in terms of accuracy. 

This thesis aims to investigate and implement in-training quantization solutions, test them with state-of-the-art datasets such as ImageNet or Visual Wake Words, and deploy the resulting model in edge devices.

Asset tracking and monitoring are becoming fundamental for e-commerce companies. Nowadays, monitoring solutions exploit standard sensors and processing techniques, thus limiting the accuracy of the result. Smart sensors such as low-cost cameras with on-board processing capabilities can improve the system’s reliability without impacting energy efficiency. Cameras can automate the picking phase by detecting objects that compose the pack and identifying damage during the transportation phase. This thesis aims to use novel object detection techniques on smart cameras (e.g., https://openmv.io/collections/cams/products/openmv-cam-h7-plus) with state-of-the-art datasets for object classification in a box.