In the banking world trades of securities are finalized every day, on behalf of the banks themselves or of their clients. When the trades have been booked by the front office the confirmations sent by the counterparty have to be checked and connected to the correct trade by hand, posing the question whether this process could not be automated using machine learning techniques. There is no straightforward solution to this problem since the confirmations differ between counterparties, and can contain different enriched information or even be in different formats. This thesis addresses the problem of matching trades with their corresponding confirmations via deep learning methods. A model is trained using contrastive learning methods on generated pairs of trades and confirmations, with the goal of matching the pairs in the latent space by using nearest neighbor classification. Accuracy is measured by dividing the correctly classified samples by the total number of samples in a testing batch. The model achieves an accuracy as high as 97.8% over 100 trade-confirmation samples with a 30-dimensional latent space, and it is shown that similar contrastive methods can indeed be used in order to solve this problem. / Banker handlar varje dag med värdpapper av olika slag, antingen för sin egen vinning eller för sina kunders. När en affär har blivit beslutad mellan två parter så bokförs denna i bägge parternas interna system. En konfirmation kommer sedan skickas från den andra parten som manuellt måste paras ihop med affären vilket väcker frågan om huruvida detta inte kan automatiseras med hjälp av maskininlärning. Det finns inte en uppenbar lösning på detta problemet då konfirmationsmeddelandena kan skiljer sig åt mellan olika parter och kan innehålla olika tillagd information eller till och med vara i olika format. En model tränas genom att använda kontrast-inlärning på genererade par av affärer och konfirmationer av affärer för att kunna para ihop paren i det latenta rummet genom att se vilka grannar som ligger närmast. Nogrannheten mäts genom att dela antalet korrekt klassificerade exempel med det totala antalet par i en grupp test-par. Modellen uppnår en noggrannhet så hög som 97.8% på 100 affärs-konfirmationspar med ett 30-dimensionellt latent rum, och det visas att kontrast-inlärning kan användas för att lösa problemet. Det är dock svårt att säga mycket om hur väl modellen kan generalisera de inlärda kunskaperna eftersom träningsdatan behövde genereras och därför saknar en del av komplexiteten av ett äkte data set.

Hyperspectral imaging is an expanding topic within the field of computer vision, that uses images of high spectral granularity. Contrastive learning is a discrim- inative approach to self-supervised learning, a form of unsupervised learning where the network is trained using self-created pseudo-labels. This work com- bines these two research areas and investigates how a pretrained network based on contrastive learning can be used for hyperspectral images. The hyperspectral images used in this work are generated from simulated RGB images and spec- tra from a spectral library. The network is trained with a pretext task based on data augmentations, and is evaluated through transfer learning and fine-tuning for a downstream task. The goal is to determine the impact of the pretext task on the downstream task and to determine the required amount of labelled data. The results show that the downstream task (a classifier) based on the pretrained network barely performs better than a classifier without a pretrained network. In the end, more research needs to be done to confirm or reject the benefit of a pretrained network based on contrastive learning for hyperspectral images. Also, the pretrained network should be tested on real-world hyperspectral data and trained with a pretext task designed for hyperspectral images.

The state-of-the-art object detection and image classification methods can perform impressively on more than 9k classes. In contrast, the number of classes in semantic segmentation datasets are fairly limited. This is not surprising , when the restrictions caused by the lack of labeled data and high computation demand are considered. To efficiently perform pixel-wise classification for c number of classes, segmentation models use cross-entropy loss on c-channel output for each pixel. The computational demand for such prediction turns out to be a major bottleneck for higher number of classes. The major goal of this thesis is to reduce the number of channels of the output prediction, thus allowing to perform semantic segmentation with very high number of classes. The reduction of dimension has been approached using metric learning for the semantic feature space. The metric learning provides us the mapping from pixel to embedding with minimal, still sufficient, number of dimensions. Our proposed approximation of groundtruth class probability for cross entropy loss helps the model to place the embeddings of same class pixels closer, reducing inter-class variabilty of clusters and increasing intra-class variability. The model also learns a prototype embedding for each class. In loss function, these class embeddings behave as positive and negative samples for pixel embeddings (anchor). We show that given a limited computational memory and resources, our approach can be used for training a segmentation model for any number of classes. We perform all experiments on one GPU and show that our approach performs similar and in some cases slightly better than deeplabv3+ baseline model for Cityscapes and ADE20K dataset. We also perform experiments to understand trade-offs in terms of memory usage, inference time and performance metrics. Our work helps in alleviating the problem of computational complexity, thus paving the way for image segmentation task with very high number of semantic classes. / De ledande djupa inlärningsmetoderna inom objektdetektion och bildklassificering kan hantera väl över 9000 klasser. Inom semantisk segmentering är däremot antalet klasser begränsat för vanliga dataset. Detta är inte förvånande då det behövs mycket annoterad data och beräkningskraft. För att effektivt kunna göra en pixelvis klassificering av c klasser, använder segmenteringsmetoder den s.k. korsentropin över c sannolikhets värden för varje pixel för att träna det djupa nätverket. Beräkningskomplexiteten från detta steg är den huvudsakliga flaskhalsen för att kunna öka antalet klasser. Det huvudsakliga målet av detta examensarbete är att minska antalet kanaler i prediktionen av nätverket för att kunna prediktera semantisk segmentering även vid ett mycket högt antal klasser. För att åstadkomma detta används metric learning för att träna slutrepresentationen av nätet. Metric learning metoden låter oss träna en representation med ett minimalt, men fortfarande tillräckligt antal dimensioner. Vi föreslår en approximation av korsentropin under träning som låter modellen placera representationer från samma klass närmare varandra, vilket reducerar interklassvarians och öka intraklarrvarians. Modellen lär sig en prototyprepresentation för varje klass. För inkärningskostnadsfunktionen ses dessa prototyper som positiva och negativa representationer. Vi visar att vår metod kan användas för att träna en segmenteringsmodell för ett godtyckligt antal klasser givet begränsade minnes- och beräkningsresurser. Alla experiment genomförs på en GPU. Vår metod åstadkommer liknande eller något bättre segmenteringsprestanda än den ursprungliga deeplabv3+ modellen på Cityscapes och ADE20K dataseten. Vi genomför också experiment för att analysera avvägningen mellan minnesanvändning, beräkningstid och segmenteringsprestanda. Vår metod minskar problemet med beräkningskomplexitet, vilket banar väg för segmentering av bilder med ett stort antal semantiska klasser.

This thesis covers the topic of image matching in a visual SLAM or SfM context. These problems are generally based on a vector representation of the keypoints of one image, called a descriptor, which we seek to map to the keypoints of another, using a similarity measure to compare the descriptors. However, it remains difficult to perform this matching successfully, especially for challenging scenes where illumination changes, occlusions, motion, textureless and similar features are present, leading to mis-matched points. In this thesis, we develop a self-supervised contrastive deep learning framework for computing robust descriptors, particularly for these challenging situations.We use the TartanAir dataset built explicitly for this task, and in which these difficult scene cases are present. Our results show that descriptor learning works, improves scores, and that our method is competitive with traditional methods such as ORB. In particular, the invariance built implicitly by training pairs of positive examples through the construction of a trajectory from a sequence of images, as well as the controlled introduction of ambiguous negative examples during training, have a real observable effect on the scores obtained. / Le présent mémoire traite du sujet de mise en correspondance entre deux images dans un contexte de SLAM visuel ou de SfM. Ces problèmes reposent généralement sur une représentation vectorielle de points saillants d’une image, appelée descripteur, et qu’on cherche à mettre en correspondance avec les points saillants d’une autre, en utilisant une mesure de similarité pour comparer les descripteurs. Cependant, il reste difficile de réaliser cette mise en correspondance avec succès, en particulier pour les scènes difficiles où des changements d’illumination, des occultations, des mouvements, des éléments sans texture, et des éléments similaires sont présents, conduisant à des mises en correspondance incorrectes. Nous développons dans ce mémoire une méthode d’apprentissage profond contrastif auto-supervisé pour calculer des descripteurs robustes, particulièrement à ces situations difficiles. Nous utilisons le jeu de données TartanAir construit explicitement pour cette tâche, et dans lequel ces cas de scènes difficiles sont présents. Nos résultats montrent que l’apprentissage de descripteurs fonctionne, améliore les scores, et que notre méthode est compétitive avec les méthodes traditionnelles telles que ORB. En particulier, l’invariance bâtie implicitement en formant des paires d’exemples positifs grâce à la construction d’une trajectoire depuis une séquence d’images, ainsi que l’introduction contrôlée d’exemples négatifs ambigus pendant l’entraînement a un réel effet observable sur les scores obtenus.

Supervised learning requires labeled data which is cumbersome to produce, making it costly and time-consuming. SimCLR is a self-supervising framework that uses data augmentations to learn without labels. This thesis investigates how well cropping and color distorting augmentations work for two datasets, MPI3D and Causal3DIdent. The representations learned are evaluated using representation similarity analysis. The data augmentations were meant to make the model learn invariant representations of the object shape in the images regarding it as content while ignoring unnecessary features and regarding them as style. As a result, 8 models were created, models A-H. A and E were trained using supervised learning as a benchmark for the remaining self-supervised models. B and C learned invariant features of style instead of learning invariant representations of shape. Model D learned invariant representations of shape. Although, it also regarded style-related factors as content. Model F, G, and H managed to learn invariant representations of shape with varying intensities while regarding the rest of the features as style. The conclusion was that models can learn invariant representations of features related to content using self-supervised learning with the chosen augmentations. However, the augmentation settings must be suitable for the dataset. / Övervakad maskininlärning kräver annoterad data, vilket är dyrt och tidskrävande att producera. SimCLR är ett självövervakande maskininlärningsramverk som använder datamodifieringar för att lära sig utan annoteringar. Detta examensarbete utvärderar hur väl beskärning och färgförvrängande datamodifieringar fungerar för två dataset, MPI3D och Causal3DIdent. De inlärda representationerna utvärderas med hjälp av representativ likhetsanalys. Syftet med examensarbetet var att få de självövervakande maskininlärningsmodellerna att lära sig oföränderliga representationer av objektet i bilderna. Meningen med datamodifieringarna var att påverka modellens lärande så att modellen tolkar objektets form som relevant innehåll, men resterande egenskaper som icke-relevant innehåll. Åtta modeller skapades (A-H). A och E tränades med övervakad inlärning och användes som riktmärke för de självövervakade modellerna. B och C lärde sig oföränderliga representationer som bör ha betraktas som irrelevant istället för att lära sig form. Modell D lärde sig oföränderliga representationer av form men också irrelevanta representationer. Modellerna F, G och H lyckades lära sig oföränderliga representationer av form med varierande intensitet, samtidigt som de resterande egenskaperna betraktades som irrelevant. Beskärning och färgförvrängande datamodifieringarna gör således att självövervakande modeller kan lära sig oföränderliga representationer av egenskaper relaterade till relevant innehåll. Specifika inställningar för datamodifieringar måste dock vara lämpliga för datasetet.

L'Imitation par observation (IPO) est un paradigme d'apprentissage qui consiste à entraîner des agents autonomes dans un processus de décision markovien (PDM) en observant les démonstrations d'un expert et sans avoir accès à ses actions. Ces démonstrations peuvent être des séquences d'états de l'environnement ou des observations visuelles brutes de l'environnement. Bien que le cadre utilisant des états à dimensions réduites ait permis d'obtenir des résultats convaincants avec des approches récentes, l'utilisation d'observations visuelles reste un défi important en IPO. Une des procédures très adoptée pour résoudre le problème d’IPO consiste à apprendre une fonction de récompense à partir des démonstrations, toutefois la nécessité d’analyser l'environnement et l'expert à partir de vidéos pour récompenser l'agent augmente la complexité du problème. Nous abordons ce problème avec une méthode basée sur la représentation des comportements de l'agent dans un espace vectoriel en utilisant des vidéos démonstratives. Notre approche exploite les techniques récentes d'apprentissage contrastif d'images et vidéos et utilise un algorithme de bootstrapping pour entraîner progressivement une fonction d'encodage de trajectoires à partir de la variation du comportement de l'agent. Simultanément, cette fonction récompense l'agent imitateur lors de l'exécution d'un algorithme d'apprentissage par renforcement. Notre méthode utilise un nombre limité de vidéos démonstratives et nous n'avons pas accès à comportement expert. Nos agents imitateurs montrent des performances convaincantes sur un ensemble de tâches de contrôle et démontrent que l'apprentissage d'une fonction de codage du comportement à partir de vidéos permet de construire une fonction de récompense efficace dans un PDM. / Imitation from observation (IfO) is a learning paradigm that consists of training autonomous agents in a Markov Decision Process (MDP) by observing an expert's demonstrations and without access to its actions. These demonstrations could be sequences of environment states or raw visual observations of the environment. Although the setting using low-dimensional states has allowed obtaining convincing results with recent approaches, the use of visual observations remains an important challenge in IfO. One of the most common procedures adopted to solve the IfO problem is to learn a reward function from the demonstrations, but the need to understand the environment and the expert's moves through videos to appropriately reward the learning agent increases the complexity of the problem. We approach this problem with a method that focuses on the representation of the agent’s behaviors in a latent space using demonstrative videos. Our approach exploits recent techniques of contrastive learning of image and video and uses a bootstrapping algorithm to progressively train a trajectory encoding function from the variation of the agent’s policy. Simultaneously, this function rewards the imitating agent through a Reinforcement Learning (RL) algorithm. Our method uses a limited number of demonstrative videos and we do not have access to any expert policy. Our imitating agents in experiments show convincing performances on a set of control tasks and demonstrate that learning a behavior encoding function from videos allows for building an efficient reward function in MDP.

In machine learning, a sub-field of computer science, a two-tower architecture model is a specialised type of neural network model that encodes paired data from different modalities (like text and images, sound and video, or proteomics and gene expression profiles) into a shared latent representation space. However, when training these models using a specific contrastive loss function, known as the multimodalinfoNCE loss, seems to often lead to a unique geometric phenomenon known as the modality gap. This gap is a clear geometric separation of the embeddings of the modalities in the joint contrastive latent space. This thesis investigates the modality gap in multimodal machine learning, specifically in two-tower neural networks trained with multimodal-infoNCE loss. We examine the adequacy of the current definition of the modality gap, the conditions under which the modality gap phenomenon manifests, and its impact on representation quality and downstream task performance. The approach to address these questions consists of a two-phase experimental strategy. Phase I involves a series of experiments, ranging from toy synthetic simulations to true multimodal machine learning with complex datasets, to explore and characterise the modality gap under varying conditions. Phase II focuses on modifying the modality gap and analysing representation quality, evaluating different loss functions and their impact on the modality gap. This methodical exploration allows us to systematically dissect the emergence and implications of the modality gap phenomenon, providing insights into its impact on downstream tasks, measured with proxy metrics based on semantic clustering in the shared latent representation space and modality-specific linear probe evaluation. Our findings reveal that the modality gap definition proposed by W. Liang et al. 2022, is insufficient. We demonstrate that similar modality gap magnitudes can exhibit varying linear separability between modality embeddings in the contrastive latent space and varying embedding topologies, indicating the need for additional metrics to capture the true essence of the gap. Furthermore, our experiments show that the temperature hyperparameter in the multimodal infoNCE loss function plays a crucial role in the emergence of the modality gap, and this effect varies with different data sets. This suggests that individual dataset characteristics significantly influence the modality gap's manifestation. A key finding is the consistent emergence of modality gaps with small temperature settings in the fixed temperature mode of the loss function and almost invariably under learned temperature mode settings, regardless of the initial temperature value. Additionally, we observe that the magnitude of the modality gap is influenced by distribution shifts, with the gap magnitude increasing progressively from the training set to the validation set, then to the test set, and finally to more distributionally shifted datasets. We discover that the choice of contrastive learning method, temperature settings, and temperature values is crucial in influencing the modality gap. However, reducing the gap does not consistently improve downstream task performance, suggesting its role may be more nuanced than previously understood. This insight indicates that the modality gap might be a geometric by-product of the learning methods rather than a critical determinant of representation quality. Our results encourage the need to reevaluate the modality gap's significance in multimodal contrastive learning, emphasising the importance of dataset characteristics and contrastive learning methodology.

<p>Nowadays an increasing number of people consume, share, and interact with information online. This results in posting and counter-posting on online media by different ideological groups on various polarized topics. Consequently, online media has become the primary platform for political and social influencers to directly interact with the citizens and share their perspectives, views, and stances with the goal of gaining support for their actions, bills, and legislation. Hence, understanding the perspectives and the influencing strategies in online media texts is important for an individual to avoid misinformation and improve trust between the general people and the influencers and the authoritative figures such as the government.</p> <p><br></p> <p>Automatically understanding the perspectives in online media is difficult because of two major challenges. Firstly, the proper grammar or mechanism to characterize the perspectives is not available. Recent studies in Natural Language Processing (NLP) have leveraged resources from social science to explain perspectives. For example, Policy Framing and Moral Foundation Theory are used for understanding how issues are framed and the moral appeal expressed in texts to gain support. However, these theories often fail to capture the nuances in perspectives and cannot generalize over all topics and events. Our research in this dissertation is one of the first studies that adapt social science theories in Natural Language Processing for understanding perspectives to the extent that they can capture differences in ideologies or stances. The second key challenge in understanding perspectives in online media texts is that annotated data is difficult to obtain to build automatic methods to detect the perspectives, that can generalize over the large corpus of online media text on different topics. To tackle this problem, in this dissertation, we used weak sources of supervision such as social network interaction of users who produce and interact with the messages, weak human interaction, or artificial few-shot data using Large Language Models. </p> <p><br></p> <p>Our insight is that various tasks such as perspectives, stances, sentiments toward entities, etc. are interdependent when characterizing online media messages. As a result, we proposed approaches that jointly model various interdependent problems such as perspectives, stances, sentiments toward entities, etc., and perform structured prediction to solve them jointly. Our research findings showed that the messaging choices and perspectives on online media in response to various real-life events and their prominence and contrast in different ideological camps can be efficiently captured using our developed methods.</p>

Multi-object tracking is a classic engineering problem wherein a system must keep track of the identities of a set of a priori unknown objects through a sequence, for example video. Perfect execution of this task would mean no spurious or missed detections or identities, neither swapped identities. To measure performance of tracking systems, the Higher Order Tracking Accuracy metric is often used, which takes into account both detection and association accuracy. Prior work in monocular vision-based multi-object tracking has integrated deep learning to various degrees, with deep learning based detectors and visual feature extractors being commonplace alongside motion models of varying complexities. These methods have historically combined the usage of position and appearance in their association stage using hand-crafted heuristics, featuring increasingly complex algorithms to achieve higher performance tracking. With an interest in simplifying tracking algorithms, we turn to the field of representation learning. Presenting a novel method using a Joint-Embedding Predictive Architecture, trained through a contrastive objective, we learn object feature embeddings initialized by detections from a pre-trained detector. The results are features that fuse both positional and visual features. Comparing the performance of our method on the complex DanceTrack and relatively simpler MOT17 datasets to that of the most performant heuristic-based alternative, Deep OC-SORT, we see a significant improvement of 66.1 HOTA compared to the 61.3 HOTA of Deep OC-SORT on DanceTrack. On MOT17, which features less complex motion and less training data, heuristics-based methods outperform the proposed and prior learned tracking methods. While the method lags behind the state of the art in complex scenes, which follows the tracking-by-attention paradigm, it presents a novel approach and brings with it a new avenue of possible research. / Spårning av multipla objekt är ett typiskt ingenjörsproblem där ett system måste hålla reda på identiteterna hos en uppsättning på förhand okända objekt genom en sekvens, till exempel video. Att perfekt utföra denna uppgift skulle innebära inga felaktiga eller missade detektioner eller identiteter, inte heller utbytta identiteter. För att mäta prestanda hos spårningssystem används ofta metriken HOTA, som tar hänsyn till både detektions- och associationsnoggrannhet. Tidigare arbete inom monokulär vision-baserad flerobjektsspårning har integrerat djupinlärning i olika grad, med detektorer baserade på djupinlärning och visuella funktionsutdragare som är vanliga tillsammans med rörelsemodeller av varierande komplexitet. Dessa metoder har historiskt kombinerat användningen av position och utseende i deras associationsfas med hjälp av handgjorda heuristiker, med alltmer komplexa algoritmer för att uppnå högre prestanda i spårningen. Med ett intresse för att förenkla spårningsalgoritmer, vänder vi oss till fältet för representationsinlärning. Vi presenterar en ny metod som använder en prediktiv arkitektur med gemensam inbäddning, tränad genom ett kontrastivt mål, där vi lär oss objekt representationer initierade av detektioner från en förtränad detektor. Resultatet är en funktion som sammansmälter både position och visuel information. När vi jämför vår metod på det komplexa DanceTrack och det relativt enklare MOT17-datasetet med det mest presterande heuristikbaserade alternativet, Deep OC-SORT, ser vi en betydande förbättring på 66,1 HOTA jämfört med 61,3 HOTA för Deep OC-SORT på DanceTrack. På MOT17, som har mindre komplex rörelse och mindre träningsdata, presterar heuristikbaserade metoder bättre än den föreslagna och tidigare lärande spårningsmetoderna. Även om metoden ligger efter den senaste utvecklingen i komplexa scener, som följer paradigm för spårning-genom-uppmärksamhet, presenterar den ett nytt tillvägagångssätt och för med sig möjligheter för ny forskning.

Avec l’augmentation de la popularité de l’IA et de l’apprentissage automatique, le nombre de participants a explosé dans les conférences AI/ML. Le grand nombre d’articles soumis et la nature évolutive des sujets constituent des défis supplémentaires pour les systèmes d’évaluation par les pairs qui sont cruciaux pour nos communautés scientifiques. Certaines conférences ont évolué vers l’automatisation de l’attribution des examinateurs pour les soumissions, le TPMS [1] étant l’un de ces systèmes existants. Actuellement, TPMS prépare des profils de chercheurs et de soumissions basés sur le contenu, afin de modéliser l’adéquation des paires examinateur-soumission. Dans ce travail, nous explorons différentes approches pour le réglage fin auto-supervisé des transformateurs BERT pour les données des documents de conférence. Nous démontrons quelques nouvelles approches des vues d’augmentation pour l’auto-supervision dans le traitement du langage naturel, qui jusqu’à présent était davantage axée sur les problèmes de vision par ordinateur. Nous utilisons ensuite ces représentations d’articles individuels pour construire un modèle d’expertise qui apprend à combiner la représentation des différents travaux publiés d’un examinateur et à prédire leur pertinence pour l’examen d’un article soumis. Au final, nous montrons que de meilleures représentations individuelles des papiers et une meilleure modélisation de l’expertise conduisent à de meilleures performances dans la tâche de prédiction de l’adéquation de l’examinateur. / With the increase in popularity of AI and Machine learning, participation numbers have exploded in AI/ML conferences. The large number of submission papers and the evolving nature of topics constitute additional challenges for peer-review systems that are crucial for our scientific communities. Some conferences have moved towards automating the reviewer assignment for submissions, TPMS [1] being one such existing system. Currently, TPMS prepares content-based profiles of researchers and submission papers, to model the suitability of reviewer-submission pairs. In this work, we explore different approaches to self-supervised fine-tuning of BERT transformers for conference papers data. We demonstrate some new approaches to augmentation views for self-supervision in natural language processing, which till now has been more focused on problems in computer vision. We then use these individual paper representations for building an expertise model which learns to combine the representation of different published works of a reviewer and predict their relevance for reviewing a submission paper. In the end, we show that better individual paper representations and expertise modeling lead to better performance on the reviewer suitability prediction task.