Continual learning for computer vision applications

One of the most visionary goals of Artificial Intelligence is to create a system able to mimic and eventually surpass the intelligence observed in biological systems including, ambitiously, the one observed in humans. The main distinctive strength of humans is their ability to build a deep understanding of the world by learning continuously and drawing from their experiences. This ability, which is found in various degrees in all intelligent biological beings, allows them to adapt and properly react to changes by incrementally expanding and refining their knowledge. Arguably, achieving this ability is one of the main goals of Artificial Intelligence and a cornerstone towards the creation of intelligent artificial agents. Modern Deep Learning approaches allowed researchers and industries to achieve great advancements towards the resolution of many long-standing problems in areas like Computer Vision and Natural Language Processing. However, while this current age of renewed interest in AI allowed for the creation of extremely useful applications, a concerningly limited effort is being directed towards the design of systems able to learn continuously. The biggest problem that hinders an AI system from learning incrementally is the catastrophic forgetting phenomenon. This phenomenon, which was discovered in the 90s, naturally occurs in Deep Learning architectures where classic learning paradigms are applied when learning incrementally from a stream of experiences. This dissertation revolves around the Continual Learning field, a sub-field of Machine Learning research that has recently made a comeback following the renewed interest in Deep Learning approaches. This work will focus on a comprehensive view of continual learning by considering algorithmic, benchmarking, and applicative aspects of this field. This dissertation will also touch on community aspects such as the design and creation of research tools aimed at supporting Continual Learning research, and the theoretical and practical aspects concerning public competitions in this field.

Supporting Scientific Research Through Machine and Deep Learning: Fluorescence Microscopy and Operational Intelligence Use Cases

Although the debate of what data science is has a long history and has not reached a complete consensus yet, Data Science can be summarized as the process of learning from data. Guided by the above vision, this thesis presents two independent data science projects developed in the scope of multidisciplinary applied research. The first part analyzes fluorescence microscopy images typically produced in life science experiments, where the objective is to count how many marked neuronal cells are present in each image. Aiming to automate the task for supporting research in the area, we propose a neural network architecture tuned specifically for this use case, cell ResUnet (c-ResUnet), and discuss the impact of alternative training strategies in overcoming particular challenges of our data. The approach provides good results in terms of both detection and counting, showing performance comparable to the interpretation of human operators. As a meaningful addition, we release the pre-trained model and the Fluorescent Neuronal Cells dataset collecting pixel-level annotations of where neuronal cells are located. In this way, we hope to help future research in the area and foster innovative methodologies for tackling similar problems. The second part deals with the problem of distributed data management in the context of LHC experiments, with a focus on supporting ATLAS operations concerning data transfer failures. In particular, we analyze error messages produced by failed transfers and propose a Machine Learning pipeline that leverages the word2vec language model and K-means clustering. This provides groups of similar errors that are presented to human operators as suggestions of potential issues to investigate. The approach is demonstrated on one full day of data, showing promising ability in understanding the message content and providing meaningful groupings, in line with previously reported incidents by human operators.

Application of Deep Learning techniques in the search for BSM Higgs bosons in the $\mu\mu$ final state in CMS

The Standard Model (SM) of particle physics predicts the existence of a Higgs field responsible for the generation of particles' mass. However, some aspects of this theory remain unsolved, supposing the presence of new physics Beyond the Standard Model (BSM) with the production of new particles at a higher energy scale compared to the current experimental limits. The search for additional Higgs bosons is, in fact, predicted by theoretical extensions of the SM including the Minimal Supersymmetry Standard Model (MSSM). In the MSSM, the Higgs sector consists of two Higgs doublets, resulting in five physical Higgs particles: two charged bosons $H^{\pm}$, two neutral scalars $h$ and $H$, and one pseudoscalar $A$. The work presented in this thesis is dedicated to the search of neutral non-Standard Model Higgs bosons decaying to two muons in the model independent MSSM scenario. Proton-proton collision data recorded by the CMS experiment at the CERN LHC at a center-of-mass energy of 13 TeV are used, corresponding to an integrated luminosity of $35.9\ \text{fb}^{-1}$. Such search is sensitive to neutral Higgs bosons produced either via gluon fusion process or in association with a $\text{b}\bar{\text{b}}$ quark pair. The extensive usage of Machine and Deep Learning techniques is a fundamental element in the discrimination between signal and background simulated events. A new network structure called parameterised Neural Network (pNN) has been implemented, replacing a whole set of single neural networks trained at a specific mass hypothesis value with a single neural network able to generalise well and interpolate in the entire mass range considered. The results of the pNN signal/background discrimination are used to set a model independent 95\% confidence level expected upper limit on the production cross section times branching ratio, for a generic $\phi$ boson decaying into a muon pair in the 130 to 1000 GeV range.

Development and characterization of deep learning techniques for neuroimaging data

Deep learning methods are extremely promising machine learning tools to analyze neuroimaging data. However, their potential use in clinical settings is limited because of the existing challenges of applying these methods to neuroimaging data. In this study, first a data leakage type caused by slice-level data split that is introduced during training and validation of a 2D CNN is surveyed and a quantitative assessment of the model’s performance overestimation is presented. Second, an interpretable, leakage-fee deep learning software written in a python language with a wide range of options has been developed to conduct both classification and regression analysis. The software was applied to the study of mild cognitive impairment (MCI) in patients with small vessel disease (SVD) using multi-parametric MRI data where the cognitive performance of 58 patients measured by five neuropsychological tests is predicted using a multi-input CNN model taking brain image and demographic data. Each of the cognitive test scores was predicted using different MRI-derived features. As MCI due to SVD has been hypothesized to be the effect of white matter damage, DTI-derived features MD and FA produced the best prediction outcome of the TMT-A score which is consistent with the existing literature. In a second study, an interpretable deep learning system aimed at 1) classifying Alzheimer disease and healthy subjects 2) examining the neural correlates of the disease that causes a cognitive decline in AD patients using CNN visualization tools and 3) highlighting the potential of interpretability techniques to capture a biased deep learning model is developed. Structural magnetic resonance imaging (MRI) data of 200 subjects was used by the proposed CNN model which was trained using a transfer learning-based approach producing a balanced accuracy of 71.6%. Brain regions in the frontal and parietal lobe showing the cerebral cortex atrophy were highlighted by the visualization tools.

Deep learning and embeddings for problems of computational biology

The development of Next Generation Sequencing promotes Biology in the Big Data era. The ever-increasing gap between proteins with known sequences and those with a complete functional annotation requires computational methods for automatic structure and functional annotation. My research has been focusing on proteins and led so far to the development of three novel tools, DeepREx, E-SNPs&GO and ISPRED-SEQ, based on Machine and Deep Learning approaches. DeepREx computes the solvent exposure of residues in a protein chain. This problem is relevant for the definition of structural constraints regarding the possible folding of the protein. DeepREx exploits Long Short-Term Memory layers to capture residue-level interactions between positions distant in the sequence, achieving state-of-the-art performances. With DeepRex, I conducted a large-scale analysis investigating the relationship between solvent exposure of a residue and its probability to be pathogenic upon mutation. E-SNPs&GO predicts the pathogenicity of a Single Residue Variation. Variations occurring on a protein sequence can have different effects, possibly leading to the onset of diseases. E-SNPs&GO exploits protein embeddings generated by two novel Protein Language Models (PLMs), as well as a new way of representing functional information coming from the Gene Ontology. The method achieves state-of-the-art performances and is extremely time-efficient when compared to traditional approaches. ISPRED-SEQ predicts the presence of Protein-Protein Interaction sites in a protein sequence. Knowing how a protein interacts with other molecules is crucial for accurate functional characterization. ISPRED-SEQ exploits a convolutional layer to parse local context after embedding the protein sequence with two novel PLMs, greatly surpassing the current state-of-the-art. All methods are published in international journals and are available as user-friendly web servers. They have been developed keeping in mind standard guidelines for FAIRness (FAIR: Findable, Accessible, Interoperable, Reusable) and are integrated into the public collection of tools provided by ELIXIR, the European infrastructure for Bioinformatics.

Applied deep learning and data science with a human-centric and data-centric approach

The term Artificial intelligence acquired a lot of baggage since its introduction and in its current incarnation is synonymous with Deep Learning. The sudden availability of data and computing resources has opened the gates to myriads of applications. Not all are created equal though, and problems might arise especially for fields not closely related to the tasks that pertain tech companies that spearheaded DL. The perspective of practitioners seems to be changing, however. Human-Centric AI emerged in the last few years as a new way of thinking DL and AI applications from the ground up, with a special attention at their relationship with humans. The goal is designing a system that can gracefully integrate in already established workflows, as in many real-world scenarios AI may not be good enough to completely replace its humans. Often this replacement may even be unneeded or undesirable. Another important perspective comes from, Andrew Ng, a DL pioneer, who recently started shifting the focus of development from “better models” towards better, and smaller, data. He defined his approach Data-Centric AI. Without downplaying the importance of pushing the state of the art in DL, we must recognize that if the goal is creating a tool for humans to use, more raw performance may not align with more utility for the final user. A Human-Centric approach is compatible with a Data-Centric one, and we find that the two overlap nicely when human expertise is used as the driving force behind data quality. This thesis documents a series of case-studies where these approaches were employed, to different extents, to guide the design and implementation of intelligent systems. We found human expertise proved crucial in improving datasets and models. The last chapter includes a slight deviation, with studies on the pandemic, still preserving the human and data centric perspective.

Real-time anomaly detection of gamma-ray bursts for the Cherenkov Telescope Array using deep learning

The Cherenkov Telescope Array (CTA) will be the next-generation ground-based observatory to study the universe in the very-high-energy domain. The observatory will rely on a Science Alert Generation (SAG) system to analyze the real-time data from the telescopes and generate science alerts. The SAG system will play a crucial role in the search and follow-up of transients from external alerts, enabling multi-wavelength and multi-messenger collaborations. It will maximize the potential for the detection of the rarest phenomena, such as gamma-ray bursts (GRBs), which are the science case for this study. This study presents an anomaly detection method based on deep learning for detecting gamma-ray burst events in real-time. The performance of the proposed method is evaluated and compared against the Li&Ma standard technique in two use cases of serendipitous discoveries and follow-up observations, using short exposure times. The method shows promising results in detecting GRBs and is flexible enough to allow real-time search for transient events on multiple time scales. The method does not assume background nor source models and doe not require a minimum number of photon counts to perform analysis, making it well-suited for real-time analysis. Future improvements involve further tests, relaxing some of the assumptions made in this study as well as post-trials correction of the detection significance. Moreover, the ability to detect other transient classes in different scenarios must be investigated for completeness. The system can be integrated within the SAG system of CTA and deployed on the onsite computing clusters. This would provide valuable insights into the method's performance in a real-world setting and be another valuable tool for discovering new transient events in real-time. Overall, this study makes a significant contribution to the field of astrophysics by demonstrating the effectiveness of deep learning-based anomaly detection techniques for real-time source detection in gamma-ray astronomy.

Identificazione e classificazione delle lesioni endocavitarie uterine dalle immagini isteroscopiche con il deep learning

Anche se l'isteroscopia con la biopsia endometriale è il gold standard nella diagnosi della patologia intracavitaria uterina, l'esperienza dell’isteroscopista è fondamentale per una diagnosi corretta. Il Deep Learning (DL) come metodica di intelligenza artificiale potrebbe essere un aiuto per superare questo limite. Sono disponibili pochi studi con risultati preliminari e mancano ricerche che valutano le prestazioni dei modelli di DL nell'identificazione delle lesioni intrauterine e il possibile aiuto derivato dai fattori clinici. Obiettivo: Sviluppare un modello di DL per identificare e classificare le patologie endocavitarie uterine dalle immagini isteroscopiche. Metodi: È stato eseguito uno studio di coorte retrospettivo osservazionale monocentrico su una serie consecutiva di casi isteroscopici di pazienti con patologia intracavitaria uterina confermata all’esame istologico eseguiti al Policlinico S. Orsola. Le immagini isteroscopiche sono state usate per costruire un modello di DL per la classificazione e l'identificazione delle lesioni intracavitarie con e senza l'aiuto di fattori clinici (età, menopausa, AUB, terapia ormonale e tamoxifene). Come risultati dello studio abbiamo calcolato le metriche diagnostiche del modello di DL nella classificazione e identificazione delle lesioni uterine intracavitarie con e senza l'aiuto dei fattori clinici. Risultati: Abbiamo esaminato 1.500 immagini provenienti da 266 casi: 186 pazienti avevano lesioni focali benigne, 25 lesioni diffuse benigne e 55 lesioni preneoplastiche/neoplastiche. Sia per quanto riguarda la classificazione che l’identificazione, le migliori prestazioni sono state raggiunte con l'aiuto dei fattori clinici, complessivamente con precision dell'80,11%, recall dell'80,11%, specificità del 90,06%, F1 score dell’80,11% e accuratezza dell’86,74% per la classificazione. Per l’identificazione abbiamo ottenuto un rilevamento complessivo dell’85,82%, precision 93,12%, recall del 91,63% ed F1 score del 92,37%. Conclusioni: Il modello DL ha ottenuto una bassa performance nell’identificazione e classificazione delle lesioni intracavitarie uterine dalle immagini isteroscopiche. Anche se la migliore performance diagnostica è stata ottenuta con l’aiuto di fattori clinici specifici, questo miglioramento è stato scarso.

Deep learning and spatial transcriptomics to investigate thyroid tumors

There are many diseases that affect the thyroid gland, and among them are carcinoma. Thyroid cancer is the most common endocrine neoplasm and the second most frequent cancer in the 0-49 age group. This thesis deals with two studies I conducted during my PhD. The first concerns the development of a Deep Learning model to be able to assist the pathologist in screening of thyroid cytology smears. This tool created in collaboration with Prof. Diciotti, affiliated with the DEI-UNIBO "Guglielmo Marconi" Department of Electrical Energy and Information Engineering, has an important clinical implication in that it allows patients to be stratified between those who should undergo surgery and those who should not. The second concerns the application of spatial transcriptomics on well-differentiated thyroid carcinomas to better understand their invasion mechanisms and thus to better comprehend which genes may be involved in the proliferation of these tumors. This project specifically was made possible through a fruitful collaboration with the Gustave Roussy Institute in Paris. Studying thyroid carcinoma deeply is essential to improve patient care, increase survival rates, and enhance the overall understanding of this prevalent cancer. It can lead to more effective prevention, early detection, and treatment strategies that benefit both patients and the healthcare system.

A numerical methodology for the multi-objective optimization of the DI Diesel engine combustion

DI Diesel engine are widely used both for industrial and automotive applications due to their durability and fuel economy. Nonetheless, increasing environmental concerns force that type of engine to comply with increasingly demanding emission limits, so that, it has become mandatory to develop a robust design methodology of the DI Diesel combustion system focused on reduction of soot and NOx simultaneously while maintaining a reasonable fuel economy. In recent years, genetic algorithms and CFD three-dimensional combustion simulations have been successfully applied to that kind of problem. However, combining GAs optimization with actual CFD three-dimensional combustion simulations can be too onerous since a large number of calculations is usually needed for the genetic algorithm to converge, resulting in a high computational cost and, thus, limiting the suitability of this method for industrial processes. In order to make the optimization process less time-consuming, CFD simulations can be more conveniently used to generate a training set for the learning process of an artificial neural network which, once correctly trained, can be used to forecast the engine outputs as a function of the design parameters during a GA optimization performing a so-called virtual optimization. In the current work, a numerical methodology for the multi-objective virtual optimization of the combustion of an automotive DI Diesel engine, which relies on artificial neural networks and genetic algorithms, was developed.

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.

Scene Mapping and Understanding by Robotic Vision

The first mechanical Automaton concept was found in a Chinese text written in the 3rd century BC, while Computer Vision was born in the late 1960s. Therefore, visual perception applied to machines (i.e. the Machine Vision) is a young and exciting alliance. When robots came in, the new field of Robotic Vision was born, and these terms began to be erroneously interchanged. In short, we can say that Machine Vision is an engineering domain, which concern the industrial use of Vision. The Robotic Vision, instead, is a research field that tries to incorporate robotics aspects in computer vision algorithms. Visual Servoing, for example, is one of the problems that cannot be solved by computer vision only. Accordingly, a large part of this work deals with boosting popular Computer Vision techniques by exploiting robotics: e.g. the use of kinematics to localize a vision sensor, mounted as the robot end-effector. The remainder of this work is dedicated to the counterparty, i.e. the use of computer vision to solve real robotic problems like grasping objects or navigate avoiding obstacles. Will be presented a brief survey about mapping data structures most widely used in robotics along with SkiMap, a novel sparse data structure created both for robotic mapping and as a general purpose 3D spatial index. Thus, several approaches to implement Object Detection and Manipulation, by exploiting the aforementioned mapping strategies, will be proposed, along with a completely new Machine Teaching facility in order to simply the training procedure of modern Deep Learning networks.

Robotic Sensing and Manipulation of Deformable Linear Objects with Learning-based methods

Nowadays robotic applications are widespread and most of the manipulation tasks are efficiently solved. However, Deformable-Objects (DOs) still represent a huge limitation for robots. The main difficulty in DOs manipulation is dealing with the shape and dynamics uncertainties, which prevents the use of model-based approaches (since they are excessively computationally complex) and makes sensory data difficult to interpret. This thesis reports the research activities aimed to address some applications in robotic manipulation and sensing of Deformable-Linear-Objects (DLOs), with particular focus to electric wires. In all the works, a significant effort was made in the study of an effective strategy for analyzing sensory signals with various machine learning algorithms. In the former part of the document, the main focus concerns the wire terminals, i.e. detection, grasping, and insertion. First, a pipeline that integrates vision and tactile sensing is developed, then further improvements are proposed for each module. A novel procedure is proposed to gather and label massive amounts of training images for object detection with minimal human intervention. Together with this strategy, we extend a generic object detector based on Convolutional-Neural-Networks for orientation prediction. The insertion task is also extended by developing a closed-loop control capable to guide the insertion of a longer and curved segment of wire through a hole, where the contact forces are estimated by means of a Recurrent-Neural-Network. In the latter part of the thesis, the interest shifts to the DLO shape. Robotic reshaping of a DLO is addressed by means of a sequence of pick-and-place primitives, while a decision making process driven by visual data learns the optimal grasping locations exploiting Deep Q-learning and finds the best releasing point. The success of the solution leverages on a reliable interpretation of the DLO shape. For this reason, further developments are made on the visual segmentation.

Machine learning tools for protein annotation: the cases of transmembrane β-barrel and myristoylated proteins

Biology is now a “Big Data Science” thanks to technological advancements allowing the characterization of the whole macromolecular content of a cell or a collection of cells. This opens interesting perspectives, but only a small portion of this data may be experimentally characterized. From this derives the demand of accurate and efficient computational tools for automatic annotation of biological molecules. This is even more true when dealing with membrane proteins, on which my research project is focused leading to the development of two machine learning-based methods: BetAware-Deep and SVMyr. BetAware-Deep is a tool for the detection and topology prediction of transmembrane beta-barrel proteins found in Gram-negative bacteria. These proteins are involved in many biological processes and primary candidates as drug targets. BetAware-Deep exploits the combination of a deep learning framework (bidirectional long short-term memory) and a probabilistic graphical model (grammatical-restrained hidden conditional random field). Moreover, it introduced a modified formulation of the hydrophobic moment, designed to include the evolutionary information. BetAware-Deep outperformed all the available methods in topology prediction and reported high scores in the detection task. Glycine myristoylation in Eukaryotes is the binding of a myristic acid on an N-terminal glycine. SVMyr is a fast method based on support vector machines designed to predict this modification in dataset of proteomic scale. It uses as input octapeptides and exploits computational scores derived from experimental examples and mean physicochemical features. SVMyr outperformed all the available methods for co-translational myristoylation prediction. In addition, it allows (as a unique feature) the prediction of post-translational myristoylation. Both the tools here described are designed having in mind best practices for the development of machine learning-based tools outlined by the bioinformatics community. Moreover, they are made available via user-friendly web servers. All this make them valuable tools for filling the gap between sequential and annotated data.

Development of advanced methods for the simulation of the reacting mixture formation in internal combustion engines with the use of machine learning algorithms

Besides increasing the share of electric and hybrid vehicles, in order to comply with more stringent environmental protection limitations, in the mid-term the auto industry must improve the efficiency of the internal combustion engine and the well to wheel efficiency of the employed fuel. To achieve this target, a deeper knowledge of the phenomena that influence the mixture formation and the chemical reactions involving new synthetic fuel components is mandatory, but complex and time intensive to perform purely by experimentation. Therefore, numerical simulations play an important role in this development process, but their use can be effective only if they can be considered accurate enough to capture these variations. The most relevant models necessary for the simulation of the reacting mixture formation and successive chemical reactions have been investigated in the present work, with a critical approach, in order to provide instruments to define the most suitable approaches also in the industrial context, which is limited by time constraints and budget evaluations. To overcome these limitations, new methodologies have been developed to conjugate detailed and simplified modelling techniques for the phenomena involving chemical reactions and mixture formation in non-traditional conditions (e.g. water injection, biofuels etc.). Thanks to the large use of machine learning and deep learning algorithms, several applications have been revised or implemented, with the target of reducing the computing time of some traditional tasks by orders of magnitude. Finally, a complete workflow leveraging these new models has been defined and used for evaluating the effects of different surrogate formulations of the same experimental fuel on a proof-of-concept GDI engine model.