Intelligent transportation systems (ITS) are essential tools for traffic planning, analysis, and forecasting that can utilize the huge amount of traffic data available nowadays. In this work, we aggregated detailed traffic flow sensor data, Waze reports, OpenStreetMap (OSM) features, and weather data, from California Bay Area for 6 months. Using that data, we studied three novel ITS applications using convolutional neural networks (CNNs) and recurrent neural networks (RNNs). The first experiment is an analysis of the relation between roadway shapes and accident occurrence, where results show that the speed limit and number of lanes are significant predictors for major accidents on highways. The second experiment presents a novel method for forecasting congestion severity using crowdsourced data only (Waze, OSM, and weather), without the need for traffic sensor data. The third experiment studies the improvement of traffic flow forecasting using accidents, number of lanes, weather, and time-related features, where results show significant performance improvements when the additional features where used.

Background. Under the current situation, distance education has rapidly become popular among students and teachers. This educational situation has changed the traditional way of teaching in the classroom. Under this kind of circumstance, students will be required to learn independently. But at the same time, it also brings some drawbacks, and teachers cannot obtain the feedback of students’ engagement in real-time. This thesis explores the feasibility of applying a lightweight model to recognize student engagement and the practicality of the model in a distance education environment. Objectives. This thesis aims to develop and apply a lightweight model based on Convolutional Neural Network(CNN) with acceptable performance to recognize the engagement of students in the environment of distance learning. Evaluate and compare the optimized model with selected original and other models in different performance metrics. Methods. This thesis uses experiments and literature review as research methods. The literature review is conducted to select effective CNN-based models for engagement recognition and feasible strategies for optimizing chosen models. These selected and optimized models are trained, tested, evaluated and compared as independent variables in the experiments. The performance of different models is used as the dependent variable. Results. Based on the literature review results, ShuffleNet v2 is selected as the most suitable CNN architecture for solving the task of engagement recognition. Inception v3 and ResNet are used as the classic CNN architecture for comparison. The attention mechanism and replace activation function are used as optimization methods for ShuffleNet v2. The pre-experiment results show that ShuffleNet v2 using the Leaky ReLU function has the highest accuracy compared with other activation functions. The experimental results show that the optimized model performs better in engagement recognition tasks than the baseline ShuffleNet v2 model, ResNet v2 and Inception v3 models. Conclusions. Through the analysis of the experiment results, the optimized ShuffleNet v2 has the best performance and is the most suitable model for real-world applications and deployments on mobile platforms.

Learning to rank is utilized in many scenarios, such as disease-gene association, information retrieval and recommender system. Improving the prediction accuracy of the ranking model is the main target of existing works. Contextual information has a significant influence in the ranking problem, and has been proved effective to increase the prediction performance of ranking models. Then we construct similarities for different types of entities that could utilize contextual information uniformly in an extensible way. Once we have the similarities constructed by contextual information, how to uti- lize them for different types of ranking models will be the task we should tackle. In this thesis, we propose four algorithms for learning to rank with contextual informa- tion. To refine the framework of matrix factorization, we propose an area under the ROC curve (AUC) loss to conquer the sparsity problem. Clustering and sampling methods are used to utilize the contextual information in the global perspective, and an objective function with the optimal solution is proposed to exploit the contex- tual information in the local perspective. Then, for the deep learning framework, we apply the graph convolutional network (GCN) on the ranking problem with the combination of matrix factorization. Contextual information is utilized to generate the input embeddings and graph kernels for the GCN. The third method in this thesis is proposed to directly exploit the contextual information for ranking. Laplacian loss is utilized to solve the ranking problem, which could optimize the ranking matrix directly. With this loss, entities with similar contextual information will have similar ranking results. Finally, we propose a two-step method to solve the ranking problem of the sequential data. The first step in this two-step method is to generate the em- beddings for all entities with a new sampling strategy. Graph neural network (GNN) and long short-term memory (LSTM) are combined to generate the representation of sequential data. Once we have the representation of the sequential data, we could solve the ranking problem of them with pair-wise loss and sampling strategy.

We trained a combined deep convolutional neural network to predict seals’ age (3 categories) and gender (2 categories). The entire dataset contains 110 seals with around 489 thousand location records. Most records are continuous and measured in a certain step. We created five convolutional layers for feature representation and established two fully connected structure as age’s and gender’s classifier, respectively. Each classifier consists of three fully connected layers. Treating seals’ latitude and longitude as input, entire deep learning network, which includes 780,000 neurons and 2,097,000 parameters, can reach to 70.72% accuracy rate for predicting seals’ age and simultaneously achieve 79.95% for gender estimation.

Witchweeds and broomrapes are root parasitic weeds that represent one of the main threats to global food security. By drastically reducing host crops’ yield, the parasites are often responsible for enormous economic losses estimated in billions of dollars annually. Parasitic plants rely on a chemical cue in the rhizosphere, indicating the presence of a host plant in proximity. Using this host dependency, research in parasitic plants focuses on understanding the necessary triggers for parasitic seeds germination, to either reduce their germination in presence of crops or provoke germination without hosts (i.e. suicidal germination). For this purpose, a number of synthetic analogs and inhibitors have been developed and their biological activities studied on parasitic plants around the world using various protocols. Current studies are using germination-based bioassays, where pre-conditioned parasitic seeds are placed in the presence of a chemical or plant root exudates, from which the germination ratio is assessed. Although these protocols are very sensitive at the chemical level, the germination rate recording is time consuming, represents a challenging task for researchers, and could easily be sped up leveraging automated seeds detection algorithms. In order to accelerate such protocols, we propose an automatic seed censing tool using computer vision latest development. We use a deep learning approach for object detection with the algorithm Faster R-CNN to count and discriminate germinated from non-germinated seeds. Our method has shown an accuracy of 95% in counting seeds on completely new images, and reduces the counting time by a significant margin, from 5 min to a fraction of second per image. We believe our proposed software 5 “SeedQuant” will be of great help for lab bioassays to perform large scale chemicals screening for parasitic seeds applications.

Neural networks optimizers are dominated by first-order methods, due to their inexpensive computational cost per iteration. However, it has been shown that firstorder optimization is prone to reaching sharp minima when trained with large batch sizes. As the batch size increases, the statistical stability of the problem increases, a regime that is well suited for second-order optimization methods. In this thesis, we study a distributed ellipsoidal trust region model for neural networks. We use a block diagonal approximation of the Hessian, assigning consecutive layers of the network to each process. We solve in parallel for the update direction of each subset of the parameters. We show that our optimizer is fit for large batch training as well as increasing number of processes.

Systems for running distributed deep learning training on the cloud have recently been developed. An important component of a distributed deep learning job handler is its resource allocation scheduler. This scheduler allocates computing resources to parts of a distributed training architecture. In this thesis, a serverless distributed deep learning job handler using Kubernetes was built to compare the job completion time when two different Kubernetes schedulers are used. The default Kubernetes scheduler and a gang-like custom scheduler. These schedulers were compared by performing experiments with different configurations of deep learning models, resource count selection and number of concurrent jobs. No significant difference in job completion time between the schedulers could be found. However, two benefits were found in the gang scheduler compared to the default scheduler. First, prevention of resource deadlocks where one or multiple jobs are locking resources but are unable to start. Second, reduced risk of epoch straggling, where jobs are allocated too few workers to be able to complete epochs in a reasonable time. Thus preventing other jobs from using the resources locked by the straggler job.

Cette thèse porte sur les problèmes de sécurité sur le réseau électrique français exploité par RTE, le Gestionnaire de Réseau de Transport (GRT). Les progrès en matière d'énergie durable, d'efficacité du marché de l'électricité ou de nouveaux modes de consommation poussent les GRT à exploiter le réseau plus près de ses limites de sécurité. Pour ce faire, il est essentiel de rendre le réseau plus "intelligent". Pour s'attaquer à ce problème, ce travail explore les avantages des réseaux neuronaux artificiels. Nous proposons de nouveaux algorithmes et architectures d'apprentissage profond pour aider les opérateurs humains (dispatcheurs) à prendre des décisions que nous appelons " guided dropout ". Ceci permet de prévoir les flux électriques consécutifs à une modification volontaire ou accidentelle du réseau. Pour se faire, les données continues (productions et consommations) sont introduites de manière standard, via une couche d'entrée au réseau neuronal, tandis que les données discrètes (topologies du réseau électrique) sont encodées directement dans l'architecture réseau neuronal. L’architecture est modifiée dynamiquement en fonction de la topologie du réseau électrique en activant ou désactivant des unités cachées. Le principal avantage de cette technique réside dans sa capacité à prédire les flux même pour des topologies de réseau inédites. Le "guided dropout" atteint une précision élevée (jusqu'à 99% de précision pour les prévisions de débit) tout en allant 300 fois plus vite que des simulateurs de grille physiques basés sur les lois de Kirchoff, même pour des topologies jamais vues, sans connaissance détaillée de la structure de la grille. Nous avons également montré que le "guided dropout" peut être utilisé pour classer par ordre de gravité des évènements pouvant survenir. Dans cette application, nous avons démontré que notre algorithme permet d'obtenir le même risque que les politiques actuellement mises en œuvre tout en n'exigeant que 2 % du budget informatique. Le classement reste pertinent, même pour des cas de réseau jamais vus auparavant, et peut être utilisé pour avoir une estimation globale de la sécurité globale du réseau électrique. / This thesis addresses problems of security in the French grid operated by RTE, the French ``Transmission System Operator'' (TSO). Progress in sustainable energy, electricity market efficiency, or novel consumption patterns push TSO's to operate the grid closer to its security limits. To this end, it is essential to make the grid ``smarter''. To tackle this issue, this work explores the benefits of artificial neural networks. We propose novel deep learning algorithms and architectures to assist the decisions of human operators (TSO dispatchers) that we called “guided dropout”. This allows the predictions on power flows following of a grid willful or accidental modification. This is tackled by separating the different inputs: continuous data (productions and consumptions) are introduced in a standard way, via a neural network input layer while discrete data (grid topologies) are encoded directly in the neural network architecture. This architecture is dynamically modified based on the power grid topology by switching on or off the activation of hidden units. The main advantage of this technique lies in its ability to predict the flows even for previously unseen grid topologies. The "guided dropout" achieves a high accuracy (up to 99% of precision for flow predictions) with a 300 times speedup compared to physical grid simulators based on Kirchoff's laws even for unseen contingencies, without detailed knowledge of the grid structure. We also showed that guided dropout can be used to rank contingencies that might occur in the order of severity. In this application, we demonstrated that our algorithm obtains the same risk as currently implemented policies while requiring only 2% of today's computational budget. The ranking remains relevant even handling grid cases never seen before, and can be used to have an overall estimation of the global security of the power grid.

Volatility modelling is a field dominated by classic Econometric methods such as the Nobel Prize winning Autoregressive conditional heteroskedasticity (ARCH) model. This paper therefore investigates if the field of Deep Learning can live up to the hype and outperform classic Econometrics in forecasting of realized volatility. By letting the Heterogeneous AutoRegressive model of Realized Volatility with multiple jump components (HAR-RV-CJ) represent the Econometric field as benchmark model, we compare its efficiency in forecasting realized volatility to four Deep Learning models. The results of the experiment show that the HAR-RV-CJ performs in line with the four Deep Learning models: Feed Forward Neural Network (FNN), Recurrent Neural Network (RNN), Long Short Term Memory network (LSTM) and Gated Recurrent Unit Network (GRU). Hence, the paper cannot conclude that the field of Deep Learning is superior to classic Econometrics in forecasting of realized volatility.

Bone age assessment can be useful in a variety of ways. It can help pediatricians predict growth, puberty entrance, identify diseases, and assess if a person lacking proper identification is a minor or not. It is a time-consuming process that is also prone to intra-observer variation, which can cause problems in many ways. This thesis attempts to improve and speed up bone age assessments by using different object detection methods to detect and segment bones anatomically important for the assessment and using these segmented bones to train deep learning models to predict bone age. A dataset consisting of 12811 X-ray hand images of persons ranging from infant age to 19 years of age was used. In the first research question, we compared the performance of three state-of-the-art object detection models: Mask R-CNN, Yolo, and RetinaNet. We chose the best performing model, Yolo, to segment all the growth plates in the phalanges of the dataset. We proceeded to train four different pre-trained models: Xception, InceptionV3, VGG19, and ResNet152, using both the segmented and unsegmented dataset and compared the performance. We achieved good results using both the unsegmented and segmented dataset, although the performance was slightly better using the unsegmented dataset. The analysis suggests that we might be able to achieve a higher accuracy using the segmented dataset by adding the detection of growth plates from the carpal bones, epiphysis, and the diaphysis. The best performing model was Xception, which achieved a mean average error of 1.007 years using the unsegmented dataset and 1.193 years using the segmented dataset. / <p>Presentationen gjordes online via Zoom. </p>