Humans are capable of building holistic representations for images at various levels, from local objects, to pairwise relations, to global structures. The interpretation of structures involves reasoning over repetition and symmetry of the objects in the image. In this paper, we present the Program-Guided Image Manipulator (PG-IM), inducing neuro-symbolic program-like representations to represent and manipulate images. Given an image, PG-IM detects repeated patterns, induces symbolic programs, and manipulates the image using a neural network that is guided by the program. PG-IM learns from a single image, exploiting its internal statistics. Despite trained only on image inpainting, PG-IM is directly capable of extrapolation and regularity editing in a unified framework. Extensive experiments show that PG-IM achieves superior performance on all the tasks.

While facial biometrics has been widely used for identification purpose, it has recently been researched as medical biometrics for a range of diseases. In this chapter, we investigate the facial landmark detection for atypical 3D facial modelling in facial palsy cases, while potentially such modelling can assist the medical diagnosis using atypical facial features. In our work, a study of landmarks localisation methods such as stacked hourglass networks is conducted and evaluated to ascertain their accuracy when presented with unseen atypical faces. The evaluation highlights that the state-of-the-art stacked hourglass architecture outperforms other traditional methods.

Machine learning and especially deep learning have garneredtremendous popularity in recent years due to their increased performanceover other methods. The availability of large amount of data has aidedin the progress of deep learning. Nevertheless, deep learning models areopaque and often seen as black boxes. Thus, there is an inherent need tomake the models interpretable, especially so in the medical domain. Inthis work, we propose a locally interpretable method, which is inspiredby one of the recent tools that has gained a lot of interest, called localinterpretable model-agnostic explanations (LIME). LIME generates singleinstance level explanation by artificially generating a dataset aroundthe instance (by randomly sampling and using perturbations) and thentraining a local linear interpretable model. One of the major issues inLIME is the instability in the generated explanation, which is caused dueto the randomly generated dataset. Another issue in these kind of localinterpretable models is the local fidelity. We propose novel modificationsto LIME by employing an autoencoder, which serves as a better weightingfunction for the local model. We perform extensive comparisons withdifferent datasets and show that our proposed method results in bothimproved stability, as well as local fidelity.

Neural Networks have been shown to be sensitive to common perturbations such as blur, Gaussian noise, rotations, etc. They are also vulnerable to some artificial malicious corruptions called adversarial examples. The adversarial examples study has recently become very popular and it sometimes even reduces the term "adversarial robustness" to the term "robustness". Yet, we do not know to what extent the adversarial robustness is related to the global robustness. Similarly, we do not know if a robustness to various common perturbations such as translations or contrast losses for instance, could help with adversarial corruptions. We intend to study the links between the robustnesses of neural networks to both perturbations. With our experiments, we provide one of the first benchmark designed to estimate the robustness of neural networks to common perturbations. We show that increasing the robustness to carefully selected common perturbations, can make neural networks more robust to unseen common perturbations. We also prove that adversarial robustness and robustness to common perturbations are independent. Our results make us believe that neural network robustness should be addressed in a broader sense.

Deep neural networks (DNNs) are shown to be promising solutions in many challenging artificial intelligence tasks, including object recognition, natural language processing, and even unmanned driving. A DNN model, generally based on statistical summarization of in-house training data, aims to predict correct output given an input encountered in the wild. In general, 100% precision is therefore impossible due to its probabilistic nature. For DNN practitioners, it is very hard, if not impossible, to figure out whether the low precision of a DNN model is an inevitable result, or caused by defects such as bad network design or improper training process. This paper aims at addressing this challenging problem. We approach with a careful categorization of the root causes of low precision. We find that the internal data flow footprints of a DNN model can provide insights to locate the root cause effectively. We then develop a tool, namely, DeepMorph (DNN Tomography) to analyze the root cause, which can instantly guide a DNN developer to improve the model. Case studies on four popular datasets show the effectiveness of DeepMorph.

Modern deep learning-based recommendation systems exploit hundreds to thousands of different categorical features, each with millions of different categories ranging from clicks to posts. To respect the natural diversity within the categorical data, embeddings map each category to a unique dense representation within an embedded space. Since each categorical feature could take on as many as tens of millions of different possible categories, the embedding tables form the primary memory bottleneck during both training and inference. We propose a novel approach for reducing the embedding size in an end-to-end fashion by exploiting complementary partitions of the category set to produce a unique embedding vector for each category without explicit definition. By storing multiple smaller embedding tables based on each complementary partition and combining embeddings from each table, we define a unique embedding for each category at smaller cost. This approach may be interpreted as using a specific fixed codebook to ensure uniqueness of each category's representation. Our experimental results demonstrate the effectiveness of our approach over the hashing trick for reducing the size of the embedding tables in terms of model loss and accuracy, while retaining a similar reduction in the number of parameters.

Deep learning algorithms often require solving a highly non-linear and nonconvex unconstrained optimization problem. Methods for solving optimization problems in large-scale machine learning, such as deep learning and deep reinforcement learning (RL), are generally restricted to the class of first-order algorithms, like stochastic gradient descent (SGD). While SGD iterates are inexpensive to compute, they have slow theoretical convergence rates. Furthermore, they require exhaustive trial-and-error to fine-tune many learning parameters. Using second-order curvature information to find search directions can help with more robust convergence for non-convex optimization problems. However, computing Hessian matrices for large-scale problems is not computationally practical. Alternatively, quasi-Newton methods construct an approximate of the Hessian matrix to build a quadratic model of the objective function. Quasi-Newton methods, like SGD, require only first-order gradient information, but they can result in superlinear convergence, which makes them attractive alternatives to SGD. The limited-memory Broyden-Fletcher-Goldfarb-Shanno (L-BFGS) approach is one of the most popular quasi-Newton methods that construct positive definite Hessian approximations. In this chapter, we propose efficient optimization methods based on L-BFGS quasi-Newton methods using line search and trust-region strategies. Our methods bridge the disparity between first- and second-order methods by using gradient information to calculate low-rank updates to Hessian approximations. We provide formal convergence analysis of these methods as well as empirical results on deep learning applications, such as image classification tasks and deep reinforcement learning on a set of ATARI 2600 video games. Our results show a robust convergence with preferred generalization characteristics as well as fast training time.

The increasing use of multiple sensors requires more efficient methods to represent and classify multi-dimensional data, since these applications produce a large amount of data, demanding modern techniques for data processing. Considering these observations, we present in this paper a new method for multi-dimensional data classification which relies on two premises: 1) multi-dimensional data are usually represented by tensors, due to benefits from multilinear algebra and the established tensor factorization methods; and 2) this kind of data can be described by a subspace lying within a vector space. Subspace representation has been consistently employed for pattern-set recognition, and its tensor representation counterpart is also available in the literature. However, traditional methods do not employ discriminative information of the tensors, which degrades the classification accuracy. In this scenario, generalized difference subspace (GDS) may provide an enhanced subspace representation by reducing data redundancy and revealing discriminative structures. Since GDS is not able to directly handle tensor data, we propose a new projection called n-mode GDS, which efficiently handles tensor data. In addition, n-mode Fisher score is introduced as a class separability index and an improved metric based on the geodesic distance is provided to measure the similarity between tensor data. To confirm the advantages of the proposed method, we address the problem of representing and classifying tensor data for gesture and action recognition. The experimental results have shown that the proposed approach outperforms methods commonly used in the literature without adopting pre-trained models or transfer learning.

Deep kernel learning refers to a Gaussian process that incorporates neural networks to improve the modelling of complex functions. We present a method that makes this approach feasible for problems where the data consists of line integral measurements of the target function. The performance is illustrated on computed tomography reconstruction examples.

One of the major challenges in training deep architectures for predictive tasks is the scarcity and cost of labeled training data. Active Learning (AL) is one way of addressing this challenge. In stream-based AL, observations are continuously made available to the learner that have to decide whether to request a label or to make a prediction. The goal is to reduce the request rate while at the same time maximize prediction performance. In previous research, reinforcement learning has been used for learning the AL request/prediction strategy. In our work, we propose to equip a reinforcement learning process with memory augmented neural networks, to enhance the one-shot capabilities. Moreover, we introduce Class Margin Sampling (CMS) as an extension of the standard margin sampling to the reinforcement learning setting. This strategy aims to reduce training time and improve sample efficiency in the training process. We evaluate the proposed method on a classification task using empirical accuracy of label predictions and percentage of label requests. The results indicates that the proposed method, by making use of the memory augmented networks and CMS in the training process, outperforms existing baselines.