Concepts and methods for efficient decentralized learning in federated settings

Deep Neural Networks (DNNs) have revolutionized a wide range of applications beyond traditional machine learning and artificial intelligence fields, e.g., computer vision, healthcare, natural language processing and others. At the same time, edge devices have become central in our society, generating an unprecedented amount of data which could be used to train data-hungry models such as DNNs. However, the potentially sensitive or confidential nature of gathered data poses privacy concerns when storing and processing them in centralized locations. To this purpose, decentralized learning decouples model training from the need of directly accessing raw data, by alternating on-device training and periodic communications. The ability of distilling knowledge from decentralized data, however, comes at the cost of facing more challenging learning settings, such as coping with heterogeneous hardware and network connectivity, statistical diversity of data, and ensuring verifiable privacy guarantees. This Thesis proposes an extensive overview of decentralized learning literature, including a novel taxonomy and a detailed description of the most relevant system-level contributions in the related literature for privacy, communication efficiency, data and system heterogeneity, and poisoning defense. Next, this Thesis presents the design of an original solution to tackle communication efficiency and system heterogeneity, and empirically evaluates it on federated settings. For communication efficiency, an original method, specifically designed for Convolutional Neural Networks, is also described and evaluated against the state-of-the-art. Furthermore, this Thesis provides an in-depth review of recently proposed methods to tackle the performance degradation introduced by data heterogeneity, followed by empirical evaluations on challenging data distributions, highlighting strengths and possible weaknesses of the considered solutions. Finally, this Thesis presents a novel perspective on the usage of Knowledge Distillation as a mean for optimizing decentralized learning systems in settings characterized by data heterogeneity or system heterogeneity. Our vision on relevant future research directions close the manuscript.

Meaningful insights: explainability techniques for black-box models on tabular data

Artificial Intelligence (AI) and Machine Learning (ML) are novel data analysis techniques providing very accurate prediction results. They are widely adopted in a variety of industries to improve efficiency and decision-making, but they are also being used to develop intelligent systems. Their success grounds upon complex mathematical models, whose decisions and rationale are usually difficult to comprehend for human users to the point of being dubbed as black-boxes. This is particularly relevant in sensitive and highly regulated domains. To mitigate and possibly solve this issue, the Explainable AI (XAI) field became prominent in recent years. XAI consists of models and techniques to enable understanding of the intricated patterns discovered by black-box models. In this thesis, we consider model-agnostic XAI techniques, which can be applied to Tabular data, with a particular focus on the Credit Scoring domain. Special attention is dedicated to the LIME framework, for which we propose several modifications to the vanilla algorithm, in particular: a pair of complementary Stability Indices that accurately measure LIME stability, and the OptiLIME policy which helps the practitioner finding the proper balance among explanations' stability and reliability. We subsequently put forward GLEAMS a model-agnostic surrogate interpretable model which requires to be trained only once, while providing both Local and Global explanations of the black-box model. GLEAMS produces feature attributions and what-if scenarios, from both dataset and model perspective. Eventually, we argue that synthetic data are an emerging trend in AI, being more and more used to train complex models instead of original data. To be able to explain the outcomes of such models, we must guarantee that synthetic data are reliable enough to be able to translate their explanations to real-world individuals. To this end we propose DAISYnt, a suite of tests to measure synthetic tabular data quality and privacy.

Dynamic Human Robot Interaction Framework Using Deep Learning and Robot Operating System (ROS): a practical approach

Trying to explain to a robot what to do is a difficult undertaking, and only specific types of people have been able to do so far, such as programmers or operators who have learned how to use controllers to communicate with a robot. My internship's goal was to create and develop a framework that would make that easier. The system uses deep learning techniques to recognize a set of hand gestures, both static and dynamic. Then, based on the gesture, it sends a command to a robot. To be as generic as feasible, the communication is implemented using Robot Operating System (ROS). Furthermore, users can add new recognizable gestures and link them to new robot actions; a finite state automaton enforces the users' input verification and correct action sequence. Finally, the users can create and utilize a macro to describe a sequence of actions performable by a robot.

Tecniche di Deep Learning per il riconoscimento di errori nel codice

Il ruolo dell’informatica è diventato chiave del funzionamento del mondo moderno, ormai sempre più in progressiva digitalizzazione di ogni singolo aspetto della vita dell’individuo. Con l’aumentare della complessità e delle dimensioni dei programmi, il rilevamento di errori diventa sempre di più un’attività difficile e che necessita l’impiego di tempo e risorse. Meccanismi di analisi del codice sorgente tradizionali sono esistiti fin dalla nascita dell’informatica stessa e il loro ruolo all’interno della catena produttiva di un team di programmatori non è mai stato cosi fondamentale come lo è tuttora. Questi meccanismi di analisi, però, non sono esenti da problematiche: il tempo di esecuzione su progetti di grandi dimensioni e la percentuale di falsi positivi possono, infatti, diventare un importante problema. Per questi motivi, meccanismi fondati su Machine Learning, e più in particolare Deep Learning, sono stati sviluppati negli ultimi anni. Questo lavoro di tesi si pone l’obbiettivo di esplorare e sviluppare un modello di Deep Learning per il riconoscimento di errori in un qualsiasi file sorgente scritto in linguaggio C e C++.

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.

Motor Unit Spike Trains Recontruction from sEMG Signals for Gesture Classification on a Low-Power Processing Platform

Hand gesture recognition based on surface electromyography (sEMG) signals is a promising approach for the development of intuitive human-machine interfaces (HMIs) in domains such as robotics and prosthetics. The sEMG signal arises from the muscles' electrical activity, and can thus be used to recognize hand gestures. The decoding from sEMG signals to actual control signals is non-trivial; typically, control systems map sEMG patterns into a set of gestures using machine learning, failing to incorporate any physiological insight. This master thesis aims at developing a bio-inspired hand gesture recognition system based on neuromuscular spike extraction rather than on simple pattern recognition. The system relies on a decomposition algorithm based on independent component analysis (ICA) that decomposes the sEMG signal into its constituent motor unit spike trains, which are then forwarded to a machine learning classifier. Since ICA does not guarantee a consistent motor unit ordering across different sessions, 3 approaches are proposed: 2 ordering criteria based on firing rate and negative entropy, and a re-calibration approach that allows the decomposition model to retain information about previous sessions. Using a multilayer perceptron (MLP), the latter approach results in an accuracy up to 99.4% in a 1-subject, 1-degree of freedom scenario. Afterwards, the decomposition and classification pipeline for inference is parallelized and profiled on the PULP platform, achieving a latency < 50 ms and an energy consumption < 1 mJ. Both the classification models tested (a support vector machine and a lightweight MLP) yielded an accuracy > 92% in a 1-subject, 5-classes (4 gestures and rest) scenario. These results prove that the proposed system is suitable for real-time execution on embedded platforms and also capable of matching the accuracy of state-of-the-art approaches, while also giving some physiological insight on the neuromuscular spikes underlying the sEMG.

Identificazione di caratteristiche rilevanti in capi di abbigliamento mediante deep-learning

Il mondo della moda è in continua e costante evoluzione, non solo dal punto di vista sociale, ma anche da quello tecnologico. Nel corso del presente elaborato si è studiata la possibilità di riconoscere e segmentare abiti presenti in una immagine utilizzando reti neurali profonde e approcci moderni. Sono state, quindi, analizzate reti quali FasterRCNN, MaskRCNN, YOLOv5, FashionPedia e Match-RCNN. In seguito si è approfondito l’addestramento delle reti neurali profonde in scenari di alta parallelizzazione e su macchine dotate di molteplici GPU al fine di ridurre i tempi di addestramento. Inoltre si è sperimentata la possibilità di creare una rete per prevedere se un determinato abito possa avere successo in futuro analizzando semplicemente dati passati e una immagine del vestito in questione. Necessaria per tali compiti è stata, inoltre, una approfondita analisi dei dataset esistenti nel mondo della moda e dei metodi per utilizzarli per l’addestramento. Il presente elaborato è stato svolto nell’ambito del progetto FA.RE.TRA. per il quale l'Università di Bologna svolge un compito di consulenza per lo studio di fattibilità su reti neurali in grado di svolgere i compiti menzionati.

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.

Self-supervised Information Retrieval basato su deep metric learning

La tesi ha lo scopo di ricercare, esaminare ed implementare un sistema di Machine Learning, un Recommendation Systems per precisione, che permetta la racommandazione di documenti di natura giuridica, i quali sono già stati analizzati e categorizzati appropriatamente, in maniera ottimale, il cui scopo sarebbe quello di accompagnare un sistema già implementato di Information Retrieval, istanziato sopra una web application, che permette di ricercare i documenti giuridici appena menzionati.

Smart sampling approaches for Decision-Focused Learning

Many real-word decision- making problems are defined based on forecast parameters: for example, one may plan an urban route by relying on traffic predictions. In these cases, the conventional approach consists in training a predictor and then solving an optimization problem. This may be problematic since mistakes made by the predictor may trick the optimizer into taking dramatically wrong decisions. Recently, the field of Decision-Focused Learning overcomes this limitation by merging the two stages at training time, so that predictions are rewarded and penalized based on their outcome in the optimization problem. There are however still significant challenges toward a widespread adoption of the method, mostly related to the limitation in terms of generality and scalability. One possible solution for dealing with the second problem is introducing a caching-based approach, to speed up the training process. This project aims to investigate these techniques, in order to reduce even more, the solver calls. For each considered method, we designed a particular smart sampling approach, based on their characteristics. In the case of the SPO method, we ended up discovering that it is only necessary to initialize the cache with only several solutions; those needed to filter the elements that we still need to properly learn. For the Blackbox method, we designed a smart sampling approach, based on inferred solutions.

Metodi di deep learning per la decodifica di segnali neurali registrati dalla corteccia posteriore parietale di macaco

Le interfacce cervello-macchina (BMIs) permettono di guidare devices esterni utilizzando segnali neurali. Le BMIs rappresentano un’importante tecnologia per tentare di ripristinare funzioni perse in patologie che interrompono il canale di comunicazione tra cervello e corpo, come malattie neurodegenerative o lesioni spinali. Di importanza chiave per il corretto funzionamento di una BCI è la decodifica dei segnali neurali per trasformarli in segnali idonei per guidare devices esterni. Negli anni sono stati implementati diversi tipi di algoritmi. Tra questi gli algoritmi di machine learning imparano a riconoscere i pattern neurali di attivazione mappando con grande efficienza l’input, possibilmente l’attività dei neuroni, con l’output, ad esempio i comandi motori per guidare una possibile protesi. Tra gli algoritmi di machine learning ci si è focalizzati sulle deep neural networks (DNN). Un problema delle DNN è l’elevato tempo di training. Questo infatti prevede il calcolo dei parametri ottimali della rete per minimizzare l’errore di predizione. Per ridurre questo problema si possono utilizzare le reti neurali convolutive (CNN), reti caratterizzate da minori parametri di addestramento rispetto ad altri tipi di DNN con maggiori parametri come le reti neurali ricorrenti (RNN). In questo elaborato è esposto uno studio esplorante l’utilizzo innovativo di CNN per la decodifica dell’attività di neuroni registrati da macaco sveglio mentre svolgeva compiti motori. La CNN risultante ha consentito di ottenere risultati comparabili allo stato dell’arte con un minor numero di parametri addestrabili. Questa caratteristica in futuro potrebbe essere chiave per l’utilizzo di questo tipo di reti all’interno di BMIs grazie ai tempi di calcolo ridotti, consentendo in tempo reale la traduzione di un segnale neurale in segnali per muovere neuroprotesi.

Synthetic generation of amyloid PET images by non-linear dimensionality reduction inversion

Privacy issues and data scarcity in PET field call for efficient methods to expand datasets via synthetic generation of new data that cannot be traced back to real patients and that are also realistic. In this thesis, machine learning techniques were applied to 1001 amyloid-beta PET images, which had undergone a diagnosis of Alzheimer’s disease: the evaluations were 540 positive, 457 negative and 4 unknown. Isomap algorithm was used as a manifold learning method to reduce the dimensions of the PET dataset; a numerical scale-free interpolation method was applied to invert the dimensionality reduction map. The interpolant was tested on the PET images via LOOCV, where the removed images were compared with the reconstructed ones with the mean SSIM index (MSSIM = 0.76 ± 0.06). The effectiveness of this measure is questioned, since it indicated slightly higher performance for a method of comparison using PCA (MSSIM = 0.79 ± 0.06), which gave clearly poor quality reconstructed images with respect to those recovered by the numerical inverse mapping. Ten synthetic PET images were generated and, after having been mixed with ten originals, were sent to a team of clinicians for the visual assessment of their realism; no significant agreements were found either between clinicians and the true image labels or among the clinicians, meaning that original and synthetic images were indistinguishable. The future perspective of this thesis points to the improvement of the amyloid-beta PET research field by increasing available data, overcoming the constraints of data acquisition and privacy issues. Potential improvements can be achieved via refinements of the manifold learning and the inverse mapping stages during the PET image analysis, by exploring different combinations in the choice of algorithm parameters and by applying other non-linear dimensionality reduction algorithms. A final prospect of this work is the search for new methods to assess image reconstruction quality.

Robotic grasping: State of the art and modular real-world implementation

Robotic Grasping is an important research topic in robotics since for robots to attain more general-purpose utility, grasping is a necessary skill, but very challenging to master. In general the robots may use their perception abilities like an image from a camera to identify grasps for a given object usually unknown. A grasp describes how a robotic end-effector need to be positioned to securely grab an object and successfully lift it without lost it, at the moment state of the arts solutions are still far behind humans. In the last 5–10 years, deep learning methods take the scene to overcome classical problem like the arduous and time-consuming approach to form a task-specific algorithm analytically. In this thesis are present the progress and the approaches in the robotic grasping field and the potential of the deep learning methods in robotic grasping. Based on that, an implementation of a Convolutional Neural Network (CNN) as a starting point for generation of a grasp pose from camera view has been implemented inside a ROS environment. The developed technologies have been integrated into a pick-and-place application for a Panda robot from Franka Emika. The application includes various features related to object detection and selection. Additionally, the features have been kept as generic as possible to allow for easy replacement or removal if needed, without losing time for improvement or new testing.

Pre-processing for noise detection in gene expression classification data

Due to the imprecise nature of biological experiments, biological data is often characterized by the presence of redundant and noisy data. This may be due to errors that occurred during data collection, such as contaminations in laboratorial samples. It is the case of gene expression data, where the equipments and tools currently used frequently produce noisy biological data. Machine Learning algorithms have been successfully used in gene expression data analysis. Although many Machine Learning algorithms can deal with noise, detecting and removing noisy instances from the training data set can help the induction of the target hypothesis. This paper evaluates the use of distance-based pre-processing techniques for noise detection in gene expression data classification problems. This evaluation analyzes the effectiveness of the techniques investigated in removing noisy data, measured by the accuracy obtained by different Machine Learning classifiers over the pre-processed data.

Gene Expression Complex Networks: Synthesis, Identification, and Analysis

Thanks to recent advances in molecular biology, allied to an ever increasing amount of experimental data, the functional state of thousands of genes can now be extracted simultaneously by using methods such as cDNA microarrays and RNA-Seq. Particularly important related investigations are the modeling and identification of gene regulatory networks from expression data sets. Such a knowledge is fundamental for many applications, such as disease treatment, therapeutic intervention strategies and drugs design, as well as for planning high-throughput new experiments. Methods have been developed for gene networks modeling and identification from expression profiles. However, an important open problem regards how to validate such approaches and its results. This work presents an objective approach for validation of gene network modeling and identification which comprises the following three main aspects: (1) Artificial Gene Networks (AGNs) model generation through theoretical models of complex networks, which is used to simulate temporal expression data; (2) a computational method for gene network identification from the simulated data, which is founded on a feature selection approach where a target gene is fixed and the expression profile is observed for all other genes in order to identify a relevant subset of predictors; and (3) validation of the identified AGN-based network through comparison with the original network. The proposed framework allows several types of AGNs to be generated and used in order to simulate temporal expression data. The results of the network identification method can then be compared to the original network in order to estimate its properties and accuracy. Some of the most important theoretical models of complex networks have been assessed: the uniformly-random Erdos-Renyi (ER), the small-world Watts-Strogatz (WS), the scale-free Barabasi-Albert (BA), and geographical networks (GG). The experimental results indicate that the inference method was sensitive to average degree k variation, decreasing its network recovery rate with the increase of k. The signal size was important for the inference method to get better accuracy in the network identification rate, presenting very good results with small expression profiles. However, the adopted inference method was not sensible to recognize distinct structures of interaction among genes, presenting a similar behavior when applied to different network topologies. In summary, the proposed framework, though simple, was adequate for the validation of the inferred networks by identifying some properties of the evaluated method, which can be extended to other inference methods.