Representation and learning of visual information for pose recognition

Recovering position from sensor information is an important problem in mobile robotics, known as localisation. Localisation requires a map or some other description of the environment to provide the robot with a context to interpret sensor data. The mobile robot system under discussion is using an artificial neural representation of position. Building a geometrical map of the environment with a single camera and artificial neural networks is difficult. Instead it would be simpler to learn position as a function of the visual input. Usually when learning images, an intermediate representation is employed. An appropriate starting point for biologically plausible image representation is the complex cells of the visual cortex, which have invariance properties that appear useful for localisation. The effectiveness for localisation of two different complex cell models are evaluated. Finally the ability of a simple neural network with single shot learning to recognise these representations and localise a robot is examined.

Deep Head Pose: Gaze-Direction Estimation in Multimodal Video

In this paper we present a convolutional neuralnetwork (CNN)-based model for human head pose estimation inlow-resolution multi-modal RGB-D data. We pose the problemas one of classification of human gazing direction. We furtherfine-tune a regressor based on the learned deep classifier. Next wecombine the two models (classification and regression) to estimateapproximate regression confidence. We present state-of-the-artresults in datasets that span the range of high-resolution humanrobot interaction (close up faces plus depth information) data tochallenging low resolution outdoor surveillance data. We buildupon our robust head-pose estimation and further introduce anew visual attention model to recover interaction with theenvironment. Using this probabilistic model, we show thatmany higher level scene understanding like human-human/sceneinteraction detection can be achieved. Our solution runs inreal-time on commercial hardware

Datasets, features, learning, and models in visual recognition

Visual recognition is a fundamental research topic in computer vision. This dissertation explores datasets, features, learning, and models used for visual recognition. In order to train visual models and evaluate different recognition algorithms, this dissertation develops an approach to collect object image datasets on web pages using an analysis of text around the image and of image appearance. This method exploits established online knowledge resources (Wikipedia pages for text; Flickr and Caltech data sets for images). The resources provide rich text and object appearance information. This dissertation describes results on two datasets. The first is Berg’s collection of 10 animal categories; on this dataset, we significantly outperform previous approaches. On an additional set of 5 categories, experimental results show the effectiveness of the method. Images are represented as features for visual recognition. This dissertation introduces a text-based image feature and demonstrates that it consistently improves performance on hard object classification problems. The feature is built using an auxiliary dataset of images annotated with tags, downloaded from the Internet. Image tags are noisy. The method obtains the text features of an unannotated image from the tags of its k-nearest neighbors in this auxiliary collection. A visual classifier presented with an object viewed under novel circumstances (say, a new viewing direction) must rely on its visual examples. This text feature may not change, because the auxiliary dataset likely contains a similar picture. While the tags associated with images are noisy, they are more stable when appearance changes. The performance of this feature is tested using PASCAL VOC 2006 and 2007 datasets. This feature performs well; it consistently improves the performance of visual object classifiers, and is particularly effective when the training dataset is small. With more and more collected training data, computational cost becomes a bottleneck, especially when training sophisticated classifiers such as kernelized SVM. This dissertation proposes a fast training algorithm called Stochastic Intersection Kernel Machine (SIKMA). This proposed training method will be useful for many vision problems, as it can produce a kernel classifier that is more accurate than a linear classifier, and can be trained on tens of thousands of examples in two minutes. It processes training examples one by one in a sequence, so memory cost is no longer the bottleneck to process large scale datasets. This dissertation applies this approach to train classifiers of Flickr groups with many group training examples. The resulting Flickr group prediction scores can be used to measure image similarity between two images. Experimental results on the Corel dataset and a PASCAL VOC dataset show the learned Flickr features perform better on image matching, retrieval, and classification than conventional visual features. Visual models are usually trained to best separate positive and negative training examples. However, when recognizing a large number of object categories, there may not be enough training examples for most objects, due to the intrinsic long-tailed distribution of objects in the real world. This dissertation proposes an approach to use comparative object similarity. The key insight is that, given a set of object categories which are similar and a set of categories which are dissimilar, a good object model should respond more strongly to examples from similar categories than to examples from dissimilar categories. This dissertation develops a regularized kernel machine algorithm to use this category dependent similarity regularization. Experiments on hundreds of categories show that our method can make significant improvement for categories with few or even no positive examples.

Development of unsupervised learning methods with applications to life sciences data

Machine Learning makes computers capable of performing tasks typically requiring human intelligence. A domain where it is having a considerable impact is the life sciences, allowing to devise new biological analysis protocols, develop patients’ treatments efficiently and faster, and reduce healthcare costs. This Thesis work presents new Machine Learning methods and pipelines for the life sciences focusing on the unsupervised field. At a methodological level, two methods are presented. The first is an “Ab Initio Local Principal Path” and it is a revised and improved version of a pre-existing algorithm in the manifold learning realm. The second contribution is an improvement over the Import Vector Domain Description (one-class learning) through the Kullback-Leibler divergence. It hybridizes kernel methods to Deep Learning obtaining a scalable solution, an improved probabilistic model, and state-of-the-art performances. Both methods are tested through several experiments, with a central focus on their relevance in life sciences. Results show that they improve the performances achieved by their previous versions. At the applicative level, two pipelines are presented. The first one is for the analysis of RNA-Seq datasets, both transcriptomic and single-cell data, and is aimed at identifying genes that may be involved in biological processes (e.g., the transition of tissues from normal to cancer). In this project, an R package is released on CRAN to make the pipeline accessible to the bioinformatic Community through high-level APIs. The second pipeline is in the drug discovery domain and is useful for identifying druggable pockets, namely regions of a protein with a high probability of accepting a small molecule (a drug). Both these pipelines achieve remarkable results. Lastly, a detour application is developed to identify the strengths/limitations of the “Principal Path” algorithm by analyzing Convolutional Neural Networks induced vector spaces. This application is conducted in the music and visual arts domains.

Sviluppo di Retrieval-based Chatbot per l'Italiano con Transformer

L’Intelligenza Artificiale negli ultimi anni sta plasmando il futuro dell’umanità in quasi tutti i settori. È già il motore principale di diverse tecnologie emergenti come i big data, la robotica e l’IoT e continuerà ad agire come innovatore tecnologico nel futuro prossimo. Le recenti scoperte e migliorie sia nel campo dell’hardware che in quello matematico hanno migliorato l’efficienza e ridotto i tempi di esecuzione dei software. È in questo contesto che sta evolvendo anche il Natural Language Processing (NLP), un ramo dell’Intelligenza Artificiale che studia il modo in cui fornire ai computer l'abilità di comprendere un testo scritto o parlato allo stesso modo in cui lo farebbe un essere umano. Le ambiguità che distinguono la lingua naturale dalle altre rendono ardui gli studi in questo settore. Molti dei recenti sviluppi algoritmici su NLP si basano su tecnologie inventate decenni fa. La ricerca in questo settore è quindi in continua evoluzione. Questa tesi si pone l'obiettivo di sviluppare la logica di una chatbot help-desk per un'azienda privata. Lo scopo è, sottoposta una domanda da parte di un utente, restituire la risposta associata presente in una collezione domande-risposte. Il problema che questa tesi affronta è sviluppare un modello di NLP in grado di comprendere il significato semantico delle domande in input, poiché esse possono essere formulate in molteplici modi, preservando il contenuto semantico a discapito della sintassi. A causa delle ridotte dimensioni del dataset italiano proprietario su cui testare il modello chatbot, sono state eseguite molteplici sperimentazioni su un ulteriore dataset italiano con task affine. Attraverso diversi approcci di addestramento, tra cui apprendimento metrico, sono state raggiunte alte accuratezze sulle più comuni metriche di valutazione, confermando le capacità del modello proposto e sviluppato.

Continual learning on nano drones

Miniaturized flying robotic platforms, called nano-drones, have the potential to revolutionize the autonomous robots industry sector thanks to their very small form factor. The nano-drones’ limited payload only allows for a sub-100mW microcontroller unit for the on-board computations. Therefore, traditional computer vision and control algorithms are too computationally expensive to be executed on board these palm-sized robots, and we are forced to rely on artificial intelligence to trade off accuracy in favor of lightweight pipelines for autonomous tasks. However, relying on deep learning exposes us to the problem of generalization since the deployment scenario of a convolutional neural network (CNN) is often composed by different visual cues and different features from those learned during training, leading to poor inference performances. Our objective is to develop and deploy and adaptation algorithm, based on the concept of latent replays, that would allow us to fine-tune a CNN to work in new and diverse deployment scenarios. To do so we start from an existing model for visual human pose estimation, called PULPFrontnet, which is used to identify the pose of a human subject in space through its 4 output variables, and we present the design of our novel adaptation algorithm, which features automatic data gathering and labeling and on-device deployment. We therefore showcase the ability of our algorithm to adapt PULP-Frontnet to new deployment scenarios, improving the R2 scores of the four network outputs, with respect to an unknown environment, from approximately [−0.2, 0.4, 0.0,−0.7] to [0.25, 0.45, 0.2, 0.1]. Finally we demonstrate how it is possible to fine-tune our neural network in real time (i.e., under 76 seconds), using the target parallel ultra-low power GAP 8 System-on-Chip on board the nano-drone, and we show how all adaptation operations can take place using less than 2mWh of energy, a small fraction of the available battery power.

The Value of Imperfection : the Wabi-Sabi Principle in Aesthetics and Learning

OER-based learning has the potential to overcome many shortcomings and problems of traditional education. It is not hampered by IP restrictions; can depend on collaborative, cumulative, iterative refinement of resources; and the digital form provides unprecedented flexibility with respect to configuration and delivery. The OER community is a progressive group of educators and learners with decades of learning research to draw from, who know that we must prepare learners for an evolving and diverse reality. Despite this OER tends to replicate the unsuccessful characteristics of traditional education. To remedy this we may need to remember the importance of imperfection, mistakes, problems, disagreement, and the incomplete for engaged learning, and relinquish our notions of perfection, acknowledging that learners learn differently and we need diverse learners. We must stretch our perceptions of quality and provide mechanisms for engaging the incredible pool of educators globally to fulfill the promise of inclusive education.

Visual SLAM for underwater vehicles using video velocity log and natural landmarks

A visual SLAM system has been implemented and optimised for real-time deployment on an AUV equipped with calibrated stereo cameras. The system incorporates a novel approach to landmark description in which landmarks are local sub maps that consist of a cloud of 3D points and their associated SIFT/SURF descriptors. Landmarks are also sparsely distributed which simplifies and accelerates data association and map updates. In addition to landmark-based localisation the system utilises visual odometry to estimate the pose of the vehicle in 6 degrees of freedom by identifying temporal matches between consecutive local sub maps and computing the motion. Both the extended Kalman filter and unscented Kalman filter have been considered for filtering the observations. The output of the filter is also smoothed using the Rauch-Tung-Striebel (RTS) method to obtain a better alignment of the sequence of local sub maps and to deliver a large-scale 3D acquisition of the surveyed area. Synthetic experiments have been performed using a simulation environment in which ray tracing is used to generate synthetic images for the stereo system

Deep generative models for facial keypoints detection

A new area of machine learning research called deep learning, has moved machine learning closer to one of its original goals: artificial intelligence and general learning algorithm. The key idea is to pretrain models in completely unsupervised way and finally they can be fine-tuned for the task at hand using supervised learning. In this thesis, a general introduction to deep learning models and algorithms are given and these methods are applied to facial keypoints detection. The task is to predict the positions of 15 keypoints on grayscale face images. Each predicted keypoint is specified by an (x,y) real-valued pair in the space of pixel indices. In experiments, we pretrained deep belief networks (DBN) and finally performed a discriminative fine-tuning. We varied the depth and size of an architecture. We tested both deterministic and sampled hidden activations and the effect of additional unlabeled data on pretraining. The experimental results show that our model provides better results than publicly available benchmarks for the dataset.

Understanding deep architectures and the effect of unsupervised pre-training

Cette thèse porte sur une classe d'algorithmes d'apprentissage appelés architectures profondes. Il existe des résultats qui indiquent que les représentations peu profondes et locales ne sont pas suffisantes pour la modélisation des fonctions comportant plusieurs facteurs de variation. Nous sommes particulièrement intéressés par ce genre de données car nous espérons qu'un agent intelligent sera en mesure d'apprendre à les modéliser automatiquement; l'hypothèse est que les architectures profondes sont mieux adaptées pour les modéliser. Les travaux de Hinton (2006) furent une véritable percée, car l'idée d'utiliser un algorithme d'apprentissage non-supervisé, les machines de Boltzmann restreintes, pour l'initialisation des poids d'un réseau de neurones supervisé a été cruciale pour entraîner l'architecture profonde la plus populaire, soit les réseaux de neurones artificiels avec des poids totalement connectés. Cette idée a été reprise et reproduite avec succès dans plusieurs contextes et avec une variété de modèles. Dans le cadre de cette thèse, nous considérons les architectures profondes comme des biais inductifs. Ces biais sont représentés non seulement par les modèles eux-mêmes, mais aussi par les méthodes d'entraînement qui sont souvent utilisés en conjonction avec ceux-ci. Nous désirons définir les raisons pour lesquelles cette classe de fonctions généralise bien, les situations auxquelles ces fonctions pourront être appliquées, ainsi que les descriptions qualitatives de telles fonctions. L'objectif de cette thèse est d'obtenir une meilleure compréhension du succès des architectures profondes. Dans le premier article, nous testons la concordance entre nos intuitions---que les réseaux profonds sont nécessaires pour mieux apprendre avec des données comportant plusieurs facteurs de variation---et les résultats empiriques. Le second article est une étude approfondie de la question: pourquoi l'apprentissage non-supervisé aide à mieux généraliser dans un réseau profond? Nous explorons et évaluons plusieurs hypothèses tentant d'élucider le fonctionnement de ces modèles. Finalement, le troisième article cherche à définir de façon qualitative les fonctions modélisées par un réseau profond. Ces visualisations facilitent l'interprétation des représentations et invariances modélisées par une architecture profonde.

On Recurrent and Deep Neural Networks

L'apprentissage profond est un domaine de recherche en forte croissance en apprentissage automatique qui est parvenu à des résultats impressionnants dans différentes tâches allant de la classification d'images à la parole, en passant par la modélisation du langage. Les réseaux de neurones récurrents, une sous-classe d'architecture profonde, s'avèrent particulièrement prometteurs. Les réseaux récurrents peuvent capter la structure temporelle dans les données. Ils ont potentiellement la capacité d'apprendre des corrélations entre des événements éloignés dans le temps et d'emmagasiner indéfiniment des informations dans leur mémoire interne. Dans ce travail, nous tentons d'abord de comprendre pourquoi la profondeur est utile. Similairement à d'autres travaux de la littérature, nos résultats démontrent que les modèles profonds peuvent être plus efficaces pour représenter certaines familles de fonctions comparativement aux modèles peu profonds. Contrairement à ces travaux, nous effectuons notre analyse théorique sur des réseaux profonds acycliques munis de fonctions d'activation linéaires par parties, puisque ce type de modèle est actuellement l'état de l'art dans différentes tâches de classification. La deuxième partie de cette thèse porte sur le processus d'apprentissage. Nous analysons quelques techniques d'optimisation proposées récemment, telles l'optimisation Hessian free, la descente de gradient naturel et la descente des sous-espaces de Krylov. Nous proposons le cadre théorique des méthodes à région de confiance généralisées et nous montrons que plusieurs de ces algorithmes développés récemment peuvent être vus dans cette perspective. Nous argumentons que certains membres de cette famille d'approches peuvent être mieux adaptés que d'autres à l'optimisation non convexe. La dernière partie de ce document se concentre sur les réseaux de neurones récurrents. Nous étudions d'abord le concept de mémoire et tentons de répondre aux questions suivantes: Les réseaux récurrents peuvent-ils démontrer une mémoire sans limite? Ce comportement peut-il être appris? Nous montrons que cela est possible si des indices sont fournis durant l'apprentissage. Ensuite, nous explorons deux problèmes spécifiques à l'entraînement des réseaux récurrents, à savoir la dissipation et l'explosion du gradient. Notre analyse se termine par une solution au problème d'explosion du gradient qui implique de borner la norme du gradient. Nous proposons également un terme de régularisation conçu spécifiquement pour réduire le problème de dissipation du gradient. Sur un ensemble de données synthétique, nous montrons empiriquement que ces mécanismes peuvent permettre aux réseaux récurrents d'apprendre de façon autonome à mémoriser des informations pour une période de temps indéfinie. Finalement, nous explorons la notion de profondeur dans les réseaux de neurones récurrents. Comparativement aux réseaux acycliques, la définition de profondeur dans les réseaux récurrents est souvent ambiguë. Nous proposons différentes façons d'ajouter de la profondeur dans les réseaux récurrents et nous évaluons empiriquement ces propositions.

Visual SLAM for underwater vehicles using video velocity log and natural landmarks

A visual SLAM system has been implemented and optimised for real-time deployment on an AUV equipped with calibrated stereo cameras. The system incorporates a novel approach to landmark description in which landmarks are local sub maps that consist of a cloud of 3D points and their associated SIFT/SURF descriptors. Landmarks are also sparsely distributed which simplifies and accelerates data association and map updates. In addition to landmark-based localisation the system utilises visual odometry to estimate the pose of the vehicle in 6 degrees of freedom by identifying temporal matches between consecutive local sub maps and computing the motion. Both the extended Kalman filter and unscented Kalman filter have been considered for filtering the observations. The output of the filter is also smoothed using the Rauch-Tung-Striebel (RTS) method to obtain a better alignment of the sequence of local sub maps and to deliver a large-scale 3D acquisition of the surveyed area. Synthetic experiments have been performed using a simulation environment in which ray tracing is used to generate synthetic images for the stereo system

e-learning in project management using simulation models: a case study based on the replication of an experiment

Current e-learning systems are increasing their importance in higher education. However, the state of the art of e-learning applications, besides the state of the practice, does not achieve the level of interactivity that current learning theories advocate. In this paper, the possibility of enhancing e-learning systems to achieve deep learning has been studied by replicating an experiment in which students had to learn basic software engineering principles. One group learned these principles using a static approach, while the other group learned the same principles using a system-dynamics-based approach, which provided interactivity and feedback. The results show that, quantitatively, the latter group achieved a better understanding of the principles; furthermore, qualitatively, they enjoyed the learning experience

Odometria visual baseada em técnicas de structure from motion

Visual Odometry is the process that estimates camera position and orientation based solely on images and in features (projections of visual landmarks present in the scene) extraced from them. With the increasing advance of Computer Vision algorithms and computer processing power, the subarea known as Structure from Motion (SFM) started to supply mathematical tools composing localization systems for robotics and Augmented Reality applications, in contrast with its initial purpose of being used in inherently offline solutions aiming 3D reconstruction and image based modelling. In that way, this work proposes a pipeline to obtain relative position featuring a previously calibrated camera as positional sensor and based entirely on models and algorithms from SFM. Techniques usually applied in camera localization systems such as Kalman filters and particle filters are not used, making unnecessary additional information like probabilistic models for camera state transition. Experiments assessing both 3D reconstruction quality and camera position estimated by the system were performed, in which image sequences captured in reallistic scenarios were processed and compared to localization data gathered from a mobile robotic platform

A self-learning simulated annealing algorithm for global optimizations of electromagnetic devices

A self-learning simulated annealing algorithm is developed by combining the characteristics of simulated annealing and domain elimination methods. The algorithm is validated by using a standard mathematical function and by optimizing the end region of a practical power transformer. The numerical results show that the CPU time required by the proposed method is about one third of that using conventional simulated annealing algorithm.