In the era of Internet of Things (IoT) and big data, collecting, processing and analyzing enormous data faces unprecedented challenges even when being stored in preprocessed form. Anomaly detection, statistically viewed as identifying outliers having low probabilities from the modelling of data distribution p(x), becomes more crucial. In this Master thesis, two (supervised and unsupervised) novel deep anomaly detection frameworks are presented which can achieve state-of-art performance on a range of datasets. Capsule net is an advanced artificial neural network, being able to encode intrinsic spatial relationship between parts and a whole. This property allows it to work as both a classifier and a deep autoencoder. Taking this advantage of CapsNet, a new anomaly detection technique named AnoCapsNet is proposed and three normality score functions are designed: prediction-probability-based (PP-based) normality score function, reconstruction-error-based (RE-based) normality score function, and a normality score function that combines prediction-probability-based and reconstruction-error-based together (named as PP+RE-based normality score function) for evaluating the "outlierness" of unseen images. The results on four datasets demonstrate that the PP-based method performs consistently well, while the RE-based approach is relatively sensitive to the similarity between labeled and unlabeled images. The PP+RE-based approach effectively takes advantages of both methods and achieves state-of-the-art results. In many situations, neither the domain of anomalous samples can be fully understood, nor the domain of the normal samples is straightforward. Thus deep generative models are more suitable than supervised methods in such cases. As a variant of variational autoencoder (VAE), beta-VAE is designed for automated discovery of interpretable factorised latent representations from raw image data in a completely unsupervised manner. The t-Distributed Stochastic Neighbor Embedding (t-SNE), an unsupervised non-linear technique primarily used for data exploration and visualizing high-dimensional data, has advantages at creating a single map that reveals local and important global structure at many different scales. Taking advantages of both disentangled representation learning (using beta-VAE as an implementation) and low-dimensional neighbor embedding (using t-SNE as an implementation), another novel anomaly detection approach named AnoDM (stands for Anomaly detection based on unsupervised Disentangled representation learning and Manifold learning) is presented. A new anomaly score function is defined by combining (1) beta-VAE's reconstruction error, and (2) latent representations' distances in the t-SNE space. This is a general framework, thus any disentangled representation learning and low-dimensional embedding techniques can be applied. AnoDM is evaluated on both image and time-series data and achieves better results than models that use just one of the two measures and other existing advanced deep learning methods.

We introduce Domain Aware Continual Zero-Shot Learning (DACZSL), the task of visually recognizing images of unseen categories in unseen domains sequentially. We created DACZSL on top of the DomainNet dataset by dividing it into a sequence of tasks, where classes are incrementally provided on seen domains during training and evaluation is conducted on unseen domains for both seen and unseen classes. We also proposed a novel Domain-Invariant CZSL Network (DIN), which outperforms state-of-the-art baseline models that we adapted to DACZSL setting. We adopt a structure-based approach to alleviate forgetting knowledge from previous tasks with a small per-task private network in addition to a global shared network. To encourage the private network to capture the domain and task-specific representation, we train our model with a novel adversarial knowledge disentanglement setting to make our global network task-invariant and domain-invariant over all the tasks. Our method also learns a class-wise learnable prompt to obtain better class-level text representation, which is used to represent side information to enable zero-shot prediction of future unseen classes. Our code and benchmarks are made available at https://zero-shot-learning.github.io/daczsl.

Deep Neural Networks have long been considered black box systems, where their interpretability is a concern when applied in safety critical systems. In this work, a novel approach of interpreting the decisions of DNNs is proposed. The approach depends on exploiting generative models and the interpretability of their latent space. Three methods for ranking features are explored, two of which depend on sensitivity analysis, and the third one depends on Random Forest model. The Random Forest model was the most successful to rank the features, given its accuracy and inherent interpretability.

Numerical methods, the collective name for numerical analysis and optimization techniques, have been widely used in the field of computer vision and deep learning. In this thesis, we investigate the algorithms of some numerical methods and their relevant applications in deep learning. These studied numerical techniques mainly include differentiable matrix power functions, differentiable eigendecomposition (ED), feasible orthogonal matrix constraints in optimization and latent semantics discovery, and physics-informed techniques for solving partial differential equations in disentangled and equivariant representation learning. We first propose two numerical solvers for the faster computation of matrix square root and its inverse. The proposed algorithms are demonstrated to have considerable speedup in practical computer vision tasks. Then we turn to resolve the main issues when integrating differentiable ED into deep learning -- backpropagation instability, slow decomposition for batched matrices, and ill-conditioned input throughout the training. Some approximation techniques are first leveraged to closely approximate the backward gradients while avoiding gradient explosion, which resolves the issue of backpropagation instability. To improve the computational efficiency of ED, we propose an efficient ED solver dedicated to small and medium batched matrices that are frequently encountered as input in deep learning. Some orthogonality techniques are also proposed to improve input conditioning. All of these techniques combine to mitigate the difficulty of applying differentiable ED in deep learning. In the last part of the thesis, we rethink some key concepts in disentangled representation learning. We first investigate the relation between disentanglement and orthogonality -- the generative models are enforced with different proposed orthogonality to show that the disentanglement performance is indeed improved. We also challenge the linear assumption of the latent traversal paths and propose to model the traversal process as dynamic spatiotemporal flows on the potential landscapes. Finally, we build probabilistic generative models of sequences that allow for novel understandings of equivariance and disentanglement. We expect our investigation could pave the way for more in-depth and impactful research at the intersection of numerical methods and deep learning.

Les progrès récents en apprentissage par renforcement ont été substantiels, mais ils dépendent souvent de l'accès à l'état. Un état est un ensemble d'informations qui fournit une description concise et complète de l'environnement, englobant tous les détails pertinents nécessaires pour que l'agent puisse prendre des décisions éclairées. Cependant, de telles données détaillées sont rarement disponibles dans les situations réelles. Les images offrent une forme de données plus réaliste et accessible, mais leur complexité pose d'importants défis dans le développement de politiques robustes et efficaces. Les méthodes d'apprentissage de représentation se sont révélées prometteuses pour améliorer l'efficacité des politiques basées sur les données de pixels. Néanmoins, les politiques peinent toujours à généraliser à de nouveaux domaines, rendant l'application de l'apprentissage par renforcement basé sur les pixels impraticable pour des scénarios du monde réel. Cela souligne le besoin urgent de s'attaquer à l'adaptation de domaine dans l'apprentissage par renforcement basé sur les pixels. Cette thèse examine le potentiel de l'apprentissage de représentation causale pour améliorer l'adaptation de domaine dans l'apprentissage par renforcement. L'idée sous-jacente est que pour que les agents s'adaptent efficacement à de nouveaux domaines, ils doivent être capables d'extraire des informations de haut niveau à partir de données brutes et de comprendre les dynamiques causales qui régulent l'environnement. Pour étudier cela, nous évaluons quatre algorithmes distincts d'apprentissage de représentation causale, chacun conçu pour capturer un niveau de structure plus détaillé dans l'espace latent, évaluant leur impact sur la performance d'adaptation de domaine. Le processus implique d'abord d'apprendre une représentation causale puis de former l'agent d'apprentissage par renforcement sur cette représentation. La performance d'adaptation de domaine de ces agents est évaluée dans deux environnements de conduite autonome : CarRacing et CARLA. Nos résultats soutiennent que l'apprentissage d'une représentation latente améliore nettement l'efficacité et la robustesse dans l'apprentissage par renforcement basé sur les pixels. De plus, ils indiquent qu'apprendre une structure causale dans l'espace latent contribue à une meilleure performance d'adaptation de domaine. Cependant, la promesse de la représentation causale pour améliorer l'adaptation de domaine est tempérée par leurs demandes computationnelles substantielles. De plus, lorsque des observations de plusieurs domaines sont disponibles, cette approche ne dépasse pas l'efficacité des méthodes plus simples. Nous avons également trouvé que les agents entraînés sur des représentations qui conservent toutes les informations de l'espace latent ont tendance à surpasser les autres, suggérant que les représentations dissociées sont préférables aux représentations invariantes. / Recent advancements in reinforcement learning have been substantial, but they often depend on access to the state. A state is a set of information that provides a concise and complete description of the environment, encompassing all relevant details necessary for the agent to make informed decisions. However, such detailed data is rarely available in real-world settings. Images present a more realistic and accessible data form, but their complexity introduces considerable challenges in developing robust and efficient policies. Representation learning methods have shown promise in enhancing the efficiency of policies based on pixel data. Nonetheless, policies continue to struggle to generalize to new domains, making the application of pixel-based reinforcement learning impractical for real-world scenarios. This highlights the urgent need to address domain adaptation in pixel-based reinforcement learning. This thesis investigates the potential of causal representation learning in improving domain adaptation in reinforcement learning. The underlying premise is that for reinforcement learning agents to adapt to new domains effectively, they must be able to extract high-level information from raw data and comprehend the causal dynamics that regulate the environment. We evaluate four distinct causal representation learning algorithms, each aimed at uncovering a more intricate level of structure within the latent space, to assess their impact on domain adaptation performance. This involves first learning a causal representation, followed by training the reinforcement learning agent on this representation. The domain adaptation performance of these agents is evaluated within two autonomous driving environments: CarRacing and CARLA. Our results support that learning a latent representation enhances efficiency and robustness in pixel-based RL. Moreover, it indicates that understanding complex causal structures in the latent space leads to improved domain adaptation performance. However, the promise of advanced causal representation in augmenting domain adaptation is tempered by its substantial computational demands. Additionally, when observations from multiple domains are available, this approach does not exceed the effectiveness of simpler methods. We also found that agents trained on representations that retain all information tend to outperform others, suggesting that disentangled representations are preferable to invariant representations.

<p dir="ltr">Several incidents report misuse of synthetic speech for impersonation attacks, spreading misinformation, and supporting financial frauds. To counter such misuse, this dissertation focuses on developing methods for speech forensics. First, we present a method to detect compressed synthetic speech. The method uses comparatively 33 times less information from compressed bit stream than used by existing methods and achieve high performance. Second, we present a transformer neural network method that uses 2D spectral representation of speech signals to detect synthetic speech. The method shows high performance on detecting both compressed and uncompressed synthetic speech. Third, we present a method using an interpretable machine learning approach known as disentangled representation learning for synthetic speech detection. Fourth, we present a method for synthetic speech attribution. It identifies the source of a speech signal. If the speech is spoken by a human, we classify it as authentic/bona fide. If the speech signal is synthetic, we identify the generation method used to create it. We examine both closed-set and open-set attribution scenarios. In a closed-set scenario, we evaluate our approach only on the speech generation methods present in the training set. In an open-set scenario, we also evaluate on methods which are not present in the training set. Fifth, we propose a multi-domain method for synthetic speech localization. It processes multi-domain features obtained from a transformer using a ResNet-style MLP. We show that with relatively less number of parameters, the proposed method performs better than existing methods. Finally, we present a new direction of research in speech forensics <i>i.e.</i>, bias and fairness of synthetic speech detectors. By bias, we refer to an action in which a detector unfairly targets a specific demographic group of individuals and falsely labels their bona fide speech as synthetic. We show that existing synthetic speech detectors are gender, age and accent biased. They also have bias against bona fide speech from people with speech impairments such as stuttering. We propose a set of augmentations that simulate stuttering in speech. We show that synthetic speech detectors trained with proposed augmentation have less bias relative to detector trained without it.</p>