On the Democratization of AI: Bridging the Gap with AutoML

Riding the wave of recent groundbreaking achievements, artificial intelligence (AI) is currently the buzzword on everybody’s lips and, allowing algorithms to learn from historical data, Machine Learning (ML) emerged as its pinnacle. The multitude of algorithms, each with unique strengths and weaknesses, highlights the absence of a universal solution and poses a challenging optimization problem. In response, automated machine learning (AutoML) navigates vast search spaces within minimal time constraints. By lowering entry barriers, AutoML emerged as promising the democratization of AI, yet facing some challenges. In data-centric AI, the discipline of systematically engineering data used to build an AI system, the challenge of configuring data pipelines is rather simple. We devise a methodology for building effective data pre-processing pipelines in supervised learning as well as a data-centric AutoML solution for unsupervised learning. In human-centric AI, many current AutoML tools were not built around the user but rather around algorithmic ideas, raising ethical and social bias concerns. We contribute by deploying AutoML tools aiming at complementing, instead of replacing, human intelligence. In particular, we provide solutions for single-objective and multi-objective optimization and showcase the challenges and potential of novel interfaces featuring large language models. Finally, there are application areas that rely on numerical simulators, often related to earth observations, they tend to be particularly high-impact and address important challenges such as climate change and crop life cycles. We commit to coupling these physical simulators with (Auto)ML solutions towards a physics-aware AI. Specifically, in precision farming, we design a smart irrigation platform that: allows real-time monitoring of soil moisture, predicts future moisture values, and estimates water demand to schedule the irrigation.

Vocabulary Transfer and Knowledge Distillation for Language Model Compression

Natural Language Processing (NLP) has seen tremendous improvements over the last few years. Transformer architectures achieved impressive results in almost any NLP task, such as Text Classification, Machine Translation, and Language Generation. As time went by, transformers continued to improve thanks to larger corpora and bigger networks, reaching hundreds of billions of parameters. Training and deploying such large models has become prohibitively expensive, such that only big high tech companies can afford to train those models. Therefore, a lot of research has been dedicated to reducing a model’s size. In this thesis, we investigate the effects of Vocabulary Transfer and Knowledge Distillation for compressing large Language Models. The goal is to combine these two methodologies to further compress models without significant loss of performance. In particular, we designed different combination strategies and conducted a series of experiments on different vertical domains (medical, legal, news) and downstream tasks (Text Classification and Named Entity Recognition). Four different methods involving Vocabulary Transfer (VIPI) with and without a Masked Language Modelling (MLM) step and with and without Knowledge Distillation are compared against a baseline that assigns random vectors to new elements of the vocabulary. Results indicate that VIPI effectively transfers information of the original vocabulary and that MLM is beneficial. It is also noted that both vocabulary transfer and knowledge distillation are orthogonal to one another and may be applied jointly. The application of knowledge distillation first before subsequently applying vocabulary transfer is recommended. Finally, model performance due to vocabulary transfer does not always show a consistent trend as the vocabulary size is reduced. Hence, the choice of vocabulary size should be empirically selected by evaluation on the downstream task similar to hyperparameter tuning.

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.

Depth Estimation in Stereo Biomedical Images via Proxy-Supervised Deep Learning

In order to estimate depth through supervised deep learning-based stereo methods, it is necessary to have access to precise ground truth depth data. While the gathering of precise labels is commonly tackled by deploying depth sensors, this is not always a viable solution. For instance, in many applications in the biomedical domain, the choice of sensors capable of sensing depth at small distances with high precision on difficult surfaces (that present non-Lambertian properties) is very limited. It is therefore necessary to find alternative techniques to gather ground truth data without having to rely on external sensors. In this thesis, two different approaches have been tested to produce supervision data for biomedical images. The first aims to obtain input stereo image pairs and disparities through simulation in a virtual environment, while the second relies on a non-learned disparity estimation algorithm in order to produce noisy disparities, which are then filtered by means of hand-crafted confidence measures to create noisy labels for a subset of pixels. Among the two, the second approach, which is referred in literature as proxy-labeling, has shown the best results and has even outperformed the non-learned disparity estimation algorithm used for supervision.

Evaluation Metrics and Transfer Learning for Sales Prediction

Sales prediction plays a huge role in modern business strategies. One of it's many use cases revolves around estimating the effects of promotions. While promotions generally have a positive effect on sales of the promoted product, they can also have a negative effect on those of other products. This phenomenon is calles sales cannibalisation. Sales cannibalisation can pose a big problem to sales forcasting algorithms. A lot of times, these algorithms focus on sales over time of a single product in a single store (a couple). This research focusses on using knowledge of a product across multiple different stores. To achieve this, we applied transfer learning on a neural model developed by Kantar Consulting to demo an approach to estimating the effect of cannibalisation. Our results show a performance increase of between 10 and 14 percent. This is a very good and desired result, and Kantar will use the approach when integrating this test method into their actual systems.

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.

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.

Tree regularized neural networks for key point analysis

This thesis contributes to the ArgMining 2021 shared task on Key Point Analysis. Key Point Analysis entails extracting and calculating the prevalence of a concise list of the most prominent talking points, from an input corpus. These talking points are usually referred to as key points. Key point analysis is divided into two subtasks: Key Point Matching, which involves assigning a matching score to each key point/argument pair, and Key Point Generation, which consists of the generation of key points. The task of Key Point Matching was approached using different models: a pretrained Sentence Transformers model and a tree-constrained Graph Neural Network were tested. The best model was the fine-tuned Sentence Transformers, which achieved a mean Average Precision score of 0.75, ranking 12 compared to other participating teams. The model was then used for the subtask of Key Point Generation using the extractive method in the selection of key point candidates and the model developed for the previous subtask to evaluate them.

Neural Radiance Fields for Relighting and View Synthesis

Neural scene representation and neural rendering are new computer vision techniques that enable the reconstruction and implicit representation of real 3D scenes from a set of 2D captured images, by fitting a deep neural network. The trained network can then be used to render novel views of the scene. A recent work in this field, Neural Radiance Fields (NeRF), presented a state-of-the-art approach, which uses a simple Multilayer Perceptron (MLP) to generate photo-realistic RGB images of a scene from arbitrary viewpoints. However, NeRF does not model any light interaction with the fitted scene; therefore, despite producing compelling results for the view synthesis task, it does not provide a solution for relighting. In this work, we propose a new architecture to enable relighting capabilities in NeRF-based representations and we introduce a new real-world dataset to train and evaluate such a model. Our method demonstrates the ability to perform realistic rendering of novel views under arbitrary lighting conditions.

Autonomous pagination for newspaper from a predefined plan

Description of the development of a product able to deliver an autonomous page construction from a predefined plan. The processes involve Machine Learning techniques for text fitting on shapes, Beam Search for associations and Deep Learning for autonomous cropping of images.

Work environments implementation for genomic reporting and analytics

A global italian pharmaceutical company has to provide two work environments that favor different needs. The environments will allow to develop solutions in a controlled, secure and at the same time in an independent manner on a state-of-the-art enterprise cloud platform. The need of developing two different environments is dictated by the needs of the working units. Indeed, the first environment is designed to facilitate the creation of application related to genomics, therefore, designed more for data-scientists. This environment is capable of consuming, producing, retrieving and incorporating data, furthermore, will support the most used programming languages for genomic applications (e.g., Python, R). The proposal was to obtain a pool of ready-togo Virtual Machines with different architectures to provide best performance based on the job that needs to be carried out. The second environment has more of a traditional trait, to obtain, via ETL (Extract-Transform-Load) process, a global datamodel, resembling a classical relational structure. It will provide major BI operations (e.g., analytics, performance measure, reports, etc.) that can be leveraged both for application analysis or for internal usage. Since, both architectures will maintain large amounts of data regarding not only pharmaceutical informations but also internal company informations, it would be possible to digest the data by reporting/ analytics tools and also apply data-mining, machine learning technologies to exploit intrinsic informations. The thesis work will introduce, proposals, implementations, descriptions of used technologies/platforms and future works of the above discussed environments.

Hierarchical Multi-Label Text Classification in a Low-Resource Setting

In this thesis we address a multi-label hierarchical text classification problem in a low-resource setting and explore different approaches to identify the best one for our case. The goal is to train a model that classifies English school exercises according to a hierarchical taxonomy with few labeled data. The experiments made in this work employ different machine learning models and text representation techniques: CatBoost with tf-idf features, classifiers based on pre-trained models (mBERT, LASER), and SetFit, a framework for few-shot text classification. SetFit proved to be the most promising approach, achieving better performance when during training only a few labeled examples per class are available. However, this thesis does not consider all the hierarchical taxonomy, but only the first two levels: to address classification with the classes at the third level further experiments should be carried out, exploring methods for zero-shot text classification, data augmentation, and strategies to exploit the hierarchical structure of the taxonomy during training.

Cyber threat intelligence: identifying hardcoded secrets in GitHub

The usage of version control systems and the capabilities of storing the source code in public platforms such as GitHub increased the number of passwords, API Keys and tokens that can be found and used causing a massive security issue for people and companies. In this project, SAP's secret scanner Credential Digger is presented. How it can scan repositories to detect hardcoded secrets and how it manages to filter out the false positives between them. Moreover, how I have implemented the Credential Digger's pre-commit hook. A performance comparison between three different implementations of the hook based on how it interacts with the Machine Learning model is presented. This project also includes how it is possible to use already detected credentials to decrease the number false positive by leveraging the similarity between leaks by using the Bucket System.

Neural Machine Translation in the Field of Law. Analysis and Evaluation against Human Reference of NMT Raw Outputs of an Atto Costitutivo of S.r.l. from Italian into German.

Artificial Intelligence (AI) is gaining ever more ground in every sphere of human life, to the point that it is now even used to pass sentences in courts. The use of AI in the field of Law is however deemed quite controversial, as it could provide more objectivity yet entail an abuse of power as well, given that bias in algorithms behind AI may cause lack of accuracy. As a product of AI, machine translation is being increasingly used in the field of Law too in order to translate laws, judgements, contracts, etc. between different languages and different legal systems. In the legal setting of Company Law, accuracy of the content and suitability of terminology play a crucial role within a translation task, as any addition or omission of content or mistranslation of terms could entail legal consequences for companies. The purpose of the present study is to first assess which neural machine translation system between DeepL and ModernMT produces a more suitable translation from Italian into German of the atto costitutivo of an Italian s.r.l. in terms of accuracy of the content and correctness of terminology, and then to assess which translation proves to be closer to a human reference translation. In order to achieve the above-mentioned aims, two human and automatic evaluations are carried out based on the MQM taxonomy and the BLEU metric. Results of both evaluations show an overall better performance delivered by ModernMT in terms of content accuracy, suitability of terminology, and closeness to a human translation. As emerged from the MQM-based evaluation, its accuracy and terminology errors account for just 8.43% (as opposed to DeepL’s 9.22%), while it obtains an overall BLEU score of 29.14 (against DeepL’s 27.02). The overall performances however show that machines still face barriers in overcoming semantic complexity, tackling polysemy, and choosing domain-specific terminology, which suggests that the discrepancy with human translation may still be remarkable.

Copy-paste data augmentation for domain transfer on traffic signs

City streets carry a lot of information that can be exploited to improve the quality of the services the citizens receive. For example, autonomous vehicles need to act accordingly to all the element that are nearby the vehicle itself, like pedestrians, traffic signs and other vehicles. It is also possible to use such information for smart city applications, for example to predict and analyze the traffic or pedestrian flows. Among all the objects that it is possible to find in a street, traffic signs are very important because of the information they carry. This information can in fact be exploited both for autonomous driving and for smart city applications. Deep learning and, more generally, machine learning models however need huge quantities to learn. Even though modern models are very good at gener- alizing, the more samples the model has, the better it can generalize between different samples. Creating these datasets organically, namely with real pictures, is a very tedious task because of the wide variety of signs available in the whole world and especially because of all the possible light, orientation conditions and con- ditions in general in which they can appear. In addition to that, it may not be easy to collect enough samples for all the possible traffic signs available, cause some of them may be very rare to find. Instead of collecting pictures manually, it is possible to exploit data aug- mentation techniques to create synthetic datasets containing the signs that are needed. Creating this data synthetically allows to control the distribution and the conditions of the signs in the datasets, improving the quality and quantity of training data that is going to be used. This thesis work is about using copy-paste data augmentation to create synthetic data for the traffic sign recognition task.