Reti Neurali Ricorrenti per la generazione di musica simbolica

Il Deep Learning ha radicalmente trasformato il mondo del Machine Learning migliorando lo stato dell'arte in diversi campi che spaziano dalla computer vision al natural language processing. Non fermandosi a problemi di classificazione, negli ultimi anni, applicazioni di tipo generativo hanno portato alla creazione di immagini realistiche e documenti letterali. Il mondo della musica non è esente da una moltitudine di esperimenti nello stesso campo, con risultati ancora acerbi ma comunque potenzialmente interessanti. In questa tesi verrà discussa l'applicazione di un di modello appartenente alla famiglia del Deep Learning per la generazione di musica simbolica.

Production Lot Detection through Synthetic Generated Data

Correctness of information gathered in production environments is an essential part of quality assurance processes in many industries, this task is often performed by human resources who visually take annotations in various steps of the production flow. Depending on the performed task the correlation between where exactly the information is gathered and what it represents is more than often lost in the process. The lack of labeled data places a great boundary on the application of deep neural networks aimed at object detection tasks, moreover supervised training of deep models requires a great amount of data to be available. Reaching an adequate large collection of labeled images through classic techniques of data annotations is an exhausting and costly task to perform, not always suitable for every scenario. A possible solution is to generate synthetic data that replicates the real one and use it to fine-tune a deep neural network trained on one or more source domains to a different target domain. The purpose of this thesis is to show a real case scenario where the provided data were both in great scarcity and missing the required annotations. Sequentially a possible approach is presented where synthetic data has been generated to address those issues while standing as a training base of deep neural networks for object detection, capable of working on images taken in production-like environments. Lastly, it compares performance on different types of synthetic data and convolutional neural networks used as backbones for the model.

Intermediate linguistic task fine-tuning on Multi-lingual models

State-of-the-art NLP systems are generally based on the assumption that the underlying models are provided with vast datasets to train on. However, especially when working in multi-lingual contexts, datasets are often scarce, thus more research should be carried out in this field. This thesis investigates the benefits of introducing an additional training step when fine-tuning NLP models, named Intermediate Training, which could be exploited to augment the data used for the training phase. The Intermediate Training step is applied by training models on NLP tasks that are not strictly related to the target task, aiming to verify if the models are able to leverage the learned knowledge of such tasks. Furthermore, in order to better analyze the synergies between different categories of NLP tasks, experimentations have been extended also to Multi-Task Training, in which the model is trained on multiple tasks at the same time.

Variational Quantum Splines: Moving Beyond Linearity for Quantum Activation Functions

Activation functions within neural networks play a crucial role in Deep Learning since they allow to learn complex and non-trivial patterns in the data. However, the ability to approximate non-linear functions is a significant limitation when implementing neural networks in a quantum computer to solve typical machine learning tasks. The main burden lies in the unitarity constraint of quantum operators, which forbids non-linearity and poses a considerable obstacle to developing such non-linear functions in a quantum setting. Nevertheless, several attempts have been made to tackle the realization of the quantum activation function in the literature. Recently, the idea of the QSplines has been proposed to approximate a non-linear activation function by implementing the quantum version of the spline functions. Yet, QSplines suffers from various drawbacks. Firstly, the final function estimation requires a post-processing step; thus, the value of the activation function is not available directly as a quantum state. Secondly, QSplines need many error-corrected qubits and a very long quantum circuits to be executed. These constraints do not allow the adoption of the QSplines on near-term quantum devices and limit their generalization capabilities. This thesis aims to overcome these limitations by leveraging hybrid quantum-classical computation. In particular, a few different methods for Variational Quantum Splines are proposed and implemented, to pave the way for the development of complete quantum activation functions and unlock the full potential of quantum neural networks in the field of quantum machine learning.

Reti Generative Avversarie nel contesto di servizi di rete sociale: un caso di studio

L’intelligenza artificiale è senza dubbio uno degli argomenti attualmente più in voga nel mondo dell’informatica, sempre in costante evoluzione ed espansione in nuovi settori. In questa elaborato progettuale viene combinato l’argomento sopracitato con il mondo dei social network, che ormai sono parte integrante della quotidianità di tutti. Viene infatti analizzato lo stato dell’arte attuale delle reti neurali, in particolare delle reti generative avversarie, e vengono esaminate le principali tipologie di social network. Su questa base, infatti, verrà realizzato un sistema di rete sociale completo nel quale una GAN sarà proprio la protagonista, sfruttando le più interessanti tecnologie attualmente disponibili. Il sistema sarà disponibile sia come applicativo per dispositivi mobile che come sito web e introdurrà elementi di gamification per aumentare l’interazione con l’utente.

Design of a wearable platform for PPG analysis with multicore low-power processor

Photoplethysmography (PPG) sensors allow for noninvasive and comfortable heart-rate (HR) monitoring, suitable for compact wearable devices. However, PPG signals collected from such devices often suffer from corruption caused by motion artifacts. This is typically addressed by combining the PPG signal with acceleration measurements from an inertial sensor. Recently, different energy-efficient deep learning approaches for heart rate estimation have been proposed. To test these new solutions, in this work, we developed a highly wearable platform (42mm x 48 mm x 1.2mm) for PPG signal acquisition and processing, based on GAP9, a parallel ultra low power system-on-chip featuring nine cores RISC-V compute cluster with neural network accelerator and 1 core RISC-V controller. The hardware platform also integrates a commercial complete Optical Biosensing Module and an ARM-Cortex M4 microcontroller unit (MCU) with Bluetooth low-energy connectivity. To demonstrate the capabilities of the system, a deep learning-based approach for PPG-based HR estimation has been deployed. Thanks to the reduced power consumption of the digital computational platform, the total power budget is just 2.67 mW providing up to 5 days of operation (105 mAh battery).

Generazione di dati sintetici di risonanza magnetica cerebrale

L’applicazione degli algoritmi di Intelligenza Artificiale (AI) al settore dell’imaging medico potrebbe apportare numerosi miglioramenti alla qualità delle cure erogate ai pazienti. Tuttavia, per poterla mettere a frutto si devono ancora superare alcuni limiti legati alla necessità di grandi quantità di immagini acquisite su pazienti reali, utili nell’addestramento degli stessi algoritmi. Il principale limite è costituito dalle norme che tutelano la privacy di dati sensibili, tra cui sono incluse le immagini mediche. La generazione di grandi dataset di immagini sintetiche, ottenute con algoritmi di Deep Learning (DL), sembra essere la soluzione a questi problemi.

Predicting corticospinal excitability from pre-TMS EEG using convolutional neural networks

The amplitude of motor evoked potentials (MEPs) elicited by transcranial magnetic stimulation (TMS) of the primary motor cortex (M1) shows a large variability from trial to trial, although MEPs are evoked by the same repeated stimulus. A multitude of factors is believed to influence MEP amplitudes, such as cortical, spinal and motor excitability state. The goal of this work is to explore to which degree the variation in MEP amplitudes can be explained by the cortical state right before the stimulation. Specifically, we analyzed a dataset acquired on eleven healthy subjects comprising, for each subject, 840 single TMS pulses applied to the left M1 during acquisition of electroencephalography (EEG) and electromyography (EMG). An interpretable convolutional neural network, named SincEEGNet, was utilized to discriminate between low- and high-corticospinal excitability trials, defined according to the MEP amplitude, using in input the pre-TMS EEG. This data-driven approach enabled considering multiple brain locations and frequency bands without any a priori selection. Post-hoc interpretation techniques were adopted to enhance interpretation by identifying the more relevant EEG features for the classification. Results show that individualized classifiers successfully discriminated between low and high M1 excitability states in all participants. Outcomes of the interpretation methods suggest the importance of the electrodes situated over the TMS stimulation site, as well as the relevance of the temporal samples of the input EEG closer to the stimulation time. This novel decoding method allows causal investigation of the cortical excitability state, which may be relevant for personalizing and increasing the efficacy of therapeutic brain-state dependent brain stimulation (for example in patients affected by Parkinson’s disease).

Valutazione delle tecniche di parallelizzazione nell'addestramento di una CNN per la diagnostica del cancro al seno.

Il cancro è un processo autosufficiente e adattivo che interagisce dinamicamente con il suo microambiente, la cui diagnosi, complessa e dispendiosa in termini di tempo e numero di specialisti impiegati, viene solitamente effettuata valutando l’imaging radiografico oppure effettuando un esame istologico. L'interpretazione di tali immagini risulta generalmente molto complessa, a questo scopo sarebbe molto utile poter addestrare un computer a comprendere tali immagini potendo di fatto affiancarsi allo specialista, senza sostituirlo, al momento della diagnosi. A questo scopo è possibile affidarsi alle tecniche di apprendimento automatico, sistema alla base dell’intelligenza artificiale (AI), le quali permettono di fatto di apprendere automaticamente la rappresentazione delle caratteristiche da immagini campione. Tali tecniche di intelligenza artificiale, hanno però bisogno, per essere addestrate, di grandi quantità di dati in cui il segnale di uscita desiderato è noto, comportando di fatto un aumento delle tempistiche di addestramento. Inoltre, in ambito sanitario, i dati sono distribuiti su più archivi, dislocati sul territorio nazionale, rendendo impossibile l’utilizzo di soluzioni centralizzate. L’obbiettivo di questa trattazione sarà cercare di trovare una soluzione a queste due problematiche, ricorrendo all’utilizzo delle tecniche di parallelizzazione. A seguito dell'introduzione dello scenario biologico e delle tecniche di diagnostica ad esso associato è presentato il percorso di creazione della rete neurale. A seguito del suo addestramento sulla GPU di una singola macchina, ottenendo un'accuratezza dell'83.94% in 5 ore 48 minuti e 43 secondi, è stata introdotto la parallelizzazione ed una sua implementazione. In conclusione, sfruttando il sistema implementato, è stata distribuita la fase di addestramento prima su due macchine e poi su tre, ottenendo una diminuzione del tempo di addestramento rispettivamente del 31.4% e del 50%.

An Embedded In-Cabin Lightweight High-Performance 3D Gaze Estimation System

Gaze estimation has gained interest in recent years for being an important cue to obtain information about the internal cognitive state of humans. Regardless of whether it is the 3D gaze vector or the point of gaze (PoG), gaze estimation has been applied in various fields, such as: human robot interaction, augmented reality, medicine, aviation and automotive. In the latter field, as part of Advanced Driver-Assistance Systems (ADAS), it allows the development of cutting-edge systems capable of mitigating road accidents by monitoring driver distraction. Gaze estimation can be also used to enhance the driving experience, for instance, autonomous driving. It also can improve comfort with augmented reality components capable of being commanded by the driver's eyes. Although, several high-performance real-time inference works already exist, just a few are capable of working with only a RGB camera on computationally constrained devices, such as a microcontroller. This work aims to develop a low-cost, efficient and high-performance embedded system capable of estimating the driver's gaze using deep learning and a RGB camera. The proposed system has achieved near-SOTA performances with about 90% less memory footprint. The capabilities to generalize in unseen environments have been evaluated through a live demonstration, where high performance and near real-time inference were obtained using a webcam and a Raspberry Pi4.

Convolutional Neural Network Architectures for Template Matching

The Neural Networks customized and tested in this thesis (WaldoNet, FlowNet and PatchNet) are a first exploration and approach to the Template Matching task. The possibilities of extension are therefore many and some are proposed below. During my thesis, I have analyzed the functioning of the classical algorithms and adapted with deep learning algorithms. The features extracted from both the template and the query images resemble the keypoints of the SIFT algorithm. Then, instead of similarity function or keypoints matching, WaldoNet and PatchNet use the convolutional layer to compare the features, while FlowNet uses the correlational layer. In addition, I have identified the major challenges of the Template Matching task (affine/non-affine transformations, intensity changes...) and solved them with a careful design of the dataset.

Mitigating non-Lambertian surfaces issues in Stereo Matching with Neural Radiance Fields

Depth estimation from images has long been regarded as a preferable alternative compared to expensive and intrusive active sensors, such as LiDAR and ToF. The topic has attracted the attention of an increasingly wide audience thanks to the great amount of application domains, such as autonomous driving, robotic navigation and 3D reconstruction. Among the various techniques employed for depth estimation, stereo matching is one of the most widespread, owing to its robustness, speed and simplicity in setup. Recent developments has been aided by the abundance of annotated stereo images, which granted to deep learning the opportunity to thrive in a research area where deep networks can reach state-of-the-art sub-pixel precision in most cases. Despite the recent findings, stereo matching still begets many open challenges, two among them being finding pixel correspondences in presence of objects that exhibits a non-Lambertian behaviour and processing high-resolution images. Recently, a novel dataset named Booster, which contains high-resolution stereo pairs featuring a large collection of labeled non-Lambertian objects, has been released. The work shown that training state-of-the-art deep neural network on such data improves the generalization capabilities of these networks also in presence of non-Lambertian surfaces. Regardless being a further step to tackle the aforementioned challenge, Booster includes a rather small number of annotated images, and thus cannot satisfy the intensive training requirements of deep learning. This thesis work aims to investigate novel view synthesis techniques to augment the Booster dataset, with ultimate goal of improving stereo matching reliability in presence of high-resolution images that displays non-Lambertian surfaces.

Intelligenza artificiale e test di Turing

Obiettivo di questo lavoro di tesi consiste nell’analizzare la domanda che il matematico e logico inglese Alan Turing propose di considerare: “Can machines think?”. Il quesito, esaminato attraverso la formulazione del gioco dell’imitazione e ormai ricordato come Test di Turing, è talmente interessante da essere divenuto uno degli argomenti più discussi nell’ambito delle scienze cognitive, della filosofia della mente e dell’informatica. In particolare è stata fondata una disciplina, chiamata intelligenza artificiale o IA, che intende studiare e comprendere se e come un sistema informatico possa essere capace di simulare una mente umana e un suo tipico comportamento. Questa tesi presenta una disamina sull’intelligenza artificiale e sul Test di Turing. Dell’IA si prenderanno in esame alcune definizioni formali della disciplina, le teorie di intelligenza artificiale debole e forte, e in particolare l’esperimento mentale della Stanza Cinese, il machine learning, il deep learning e le reti neurali, alcuni loro esempi di implementazione in diversi ambiti e infine alcune questioni etiche relative all’IA. Successivamente verranno esaminati la descrizione del gioco dell’imitazione, le più importanti critiche ed obiezioni al test di Turing, una variante del test, chiamata Test di Turing Totale, il premio Loebner, le previsioni fatte dal matematico e alcuni tentativi di superamento del test, tra cui l’implementazione dei chatterbot ELIZA, ALICE ed Eugene Goostman. Saranno infine proposte delle conclusioni in merito al lavoro svolto.

Graph-based Vision-Language Model for Multi-Modal Retrieval of Fashion Products

Artificial Intelligence is reshaping the field of fashion industry in different ways. E-commerce retailers exploit their data through AI to enhance their search engines, make outfit suggestions and forecast the success of a specific fashion product. However, it is a challenging endeavour as the data they possess is huge, complex and multi-modal. The most common way to search for fashion products online is by matching keywords with phrases in the product's description which are often cluttered, inadequate and differ across collections and sellers. A customer may also browse an online store's taxonomy, although this is time-consuming and doesn't guarantee relevant items. With the advent of Deep Learning architectures, particularly Vision-Language models, ad-hoc solutions have been proposed to model both the product image and description to solve this problems. However, the suggested solutions do not exploit effectively the semantic or syntactic information of these modalities, and the unique qualities and relations of clothing items. In this work of thesis, a novel approach is proposed to address this issues, which aims to model and process images and text descriptions as graphs in order to exploit the relations inside and between each modality and employs specific techniques to extract syntactic and semantic information. The results obtained show promising performances on different tasks when compared to the present state-of-the-art deep learning architectures.

‘Data Augmentation’ per reti neurali profonde per la classificazione del segnale EEG

I recenti sviluppi nel campo dell’intelligenza artificiale hanno permesso una più adeguata classificazione del segnale EEG. Negli ultimi anni è stato dimostrato come sia possibile ottenere ottime performance di classificazione impiegando tecniche di Machine Learning (ML) e di Deep Learning (DL), facendo uso, per quest’ultime, di reti neurali convoluzionali (Convolutional Neural Networks, CNN). In particolare, il Deep Learning richiede molti dati di training mentre spesso i dataset per EEG sono limitati ed è difficile quindi raggiungere prestazioni elevate. I metodi di Data Augmentation possono alleviare questo problema. Partendo da dati reali, questa tecnica permette, la creazione di dati artificiali fondamentali per aumentare le dimensioni del dataset di partenza. L’applicazione più comune è quella di utilizzare i Data Augmentation per aumentare le dimensioni del training set, in modo da addestrare il modello/rete neurale su un numero di campioni più esteso, riducendo gli errori di classificazione. Partendo da questa idea, i Data Augmentation sono stati applicati in molteplici campi e in particolare per la classificazione del segnale EEG. In questo elaborato di tesi, inizialmente, vengono descritti metodi di Data Augmentation implementati nel corso degli anni, utilizzabili anche nell’ambito di applicazioni EEG. Successivamente, si presentano alcuni studi specifici che applicano metodi di Data Augmentation per migliorare le presentazioni di classificatori basati su EEG per l’identificazione dello stato sonno/veglia, per il riconoscimento delle emozioni, e per la classificazione di immaginazione motoria.