We present a deep-learning network that detects multiple small objects (hundreds to thousands) in a scene while simultaneously estimating their x,y pixel locations together with a characteristic feature-set (for instance, target orientation and color). All estimations are performed in a single, forward pass which makes implementing the network fast and efficient. In this paper, we describe the architecture of our network --- nicknamed ALIEN --- and detail its performance when applied to vehicle detection.

Chronic Pulmonary Aspergillosis (CPA) is a complex lung disease caused by infection with Aspergillus. Computed tomography (CT) images are frequently requested in patients with suspected and established disease, but the radiological signs on CT are difficult to quantify making accurate follow-up challenging. We propose a novel method to train Convolutional Neural Networks using only regional labels on the presence of pathological signs, to not only detect CPA, but also spatially localize pathological signs. We use average intensity projections within different ranges of Hounsfield-unit (HU) values, transforming input 3D CT scans into 2D RGB-like images. CNN architectures are trained for hierarchical tasks, leading to precise activation maps of pathological patterns. Results on a cohort of 352 subjects demonstrate high classification accuracy, localization precision and predictive power of 2 year survival. Such tool opens the way to CPA patient stratification and quantitative follow-up of CPA pathological signs, for patients under drug therapy.

Conventional optimization based methods have utilized forward models with image priors to solve inverse problems in image processing. Recently, deep neural networks (DNN) have been investigated to significantly improve the image quality of the solution for inverse problems. Most DNN based inverse problems have focused on using data-driven image priors with massive amount of data. However, these methods often do not inherit nice properties of conventional approaches using theoretically well-grounded optimization algorithms such as monotone, global convergence. Here we investigate another possibility of using DNN for inverse problems in image processing. We propose methods to use DNNs to seamlessly speed up convergence rates of conventional optimization based methods. Our DNN-incorporated scaled gradient projection methods, without breaking theoretical properties, significantly improved convergence speed over state-of-the-art conventional optimization methods such as ISTA or FISTA in practice for inverse problems such as image inpainting, compressive image recovery with partial Fourier samples, image deblurring, and medical image reconstruction with sparse-view projections.

The recently proposed Unbiased Online Recurrent Optimization (uoro) algorithm (Tallec and Ollivier, 2018) uses an unbiased approximation of rtrl to achieve fully online gradientbased learningin rnns. In this work we analyze the variance of the gradient estimate computed by uoro, and propose several possible changes to the method which reduce this variance both in theory and practice. We also contribute significantly to the theoretical and intuitive understanding of uoro (and its existing variance reduction technique), and demonstrate a fundamental connection between its gradient estimate and the one that would be computed by reinforce if small amounts of noise were added to the rnn's hidden units.

This paper studies the problem of stance detection which aims to predict the perspective (or stance) of a given document with respect to a given claim. Stance detection is a major component of automated fact checking. As annotating stances in different domains is a tedious and costly task, automatic methods based on machine learning are viable alternatives. In this paper, we focus on adversarial domain adaptation for stance detection where we assume there exists sufficient labeled data in the source domain and limited labeled data in the target domain. Extensive experiments on publicly available datasets show the effectiveness of our domain adaption model in transferring knowledge for accurate stance detection across domains.

DiffEqFlux.jl is a library for fusing neural networks and differential equations. In this work we describe differential equations from the viewpoint of data science and discuss the complementary nature between machine learning models and differential equations. We demonstrate the ability to incorporate DifferentialEquations.jl-defined differential equation problems into a Flux-defined neural network, and vice versa. The advantages of being able to use the entire DifferentialEquations.jl suite for this purpose is demonstrated by counter examples where simple integration strategies fail, but the sophisticated integration strategies provided by the DifferentialEquations.jl library succeed. This is followed by a demonstration of delay differential equations and stochastic differential equations inside of neural networks. We show high-level functionality for defining neural ordinary differential equations (neural networks embedded into the differential equation) and describe the extra models in the Flux model zoo which includes neural stochastic differential equations. We conclude by discussing the various adjoint methods used for backpropogation of the differential equation solvers. DiffEqFlux.jl is an important contribution to the area, as it allows the full weight of the differential equation solvers developed from decades of research in the scientific computing field to be readily applied to the challenges posed by machine learning and data science.

Then, a PLDA model is trained to measure thesimilarity of i-vectors. Replacing traditional i-vectors with speaker embedding models based on deep neural networks haslead to improvement in SV [4, 3]. Nonetheless, a PLDA classifier is still needed to compare the similarity of embeddings. More recently, end-to-end training of an embedding networkthat makes decision by comparing distance in the embedding to a cross-validated threshold outperformed traditional methods. For detailed comparison between embedding networksand i-vector based methods, we refer the reader to [6, 4, 3]. Building on top of these studies, our work focuses on the comparison between two different approaches for deep metric learning (TL [5, 6, 7, 8] and PNL [10]) for end-to-end speaker embedding models. Deep metric learning: End-to-end speaker embedding models can be seen as a form of deep metric learning, which has been widely studied in the machine learning literature. Early examples of metric learning with neural networks include signature[11] and face verification [12]. Both compare pairs of examples with standard similarity functions (e.g.

The loss function of deep networks is known to be non-convex but the precise nature of this nonconvexity is still an active area of research. In this work, we study the loss landscape of deep networks through the eigendecompositions of their Hessian matrix. In particular, we examine how important the negative eigenvalues are and the benefits one can observe in handling them appropriately.

A fundamental question in deep learning concerns the role played by individual layers in a deep neural network (DNN) and the transferable properties of the data representations which they learn. To the extent that layers have clear roles one should be able to optimize them separately using layer-wise loss functions. Such loss functions would describe what is the set of good data representations at each depth of the network and provide a target for layer-wise greedy optimization (LEGO). Here we introduce the Deep Gaussian Layer-wise loss functions (DGLs) which, we believe, are the first supervised layer-wise loss functions which are both explicit and competitive in terms of accuracy. The DGLs have a solid theoretical foundation, they become exact for wide DNNs, and we find that they can monitor standard end-to-end training. Being highly structured and symmetric, the DGLs provide a promising analytic route to understanding the internal representations generated by DNNs.

We propose a new attribution method for neural networks developed using first principles of causality (to the best of our knowledge, the first such). The neural network architecture is viewed as a Structural Causal Model, and a methodology to compute the causal effect of each feature on the output is presented. With reasonable assumptions on the causal structure of the input data, we propose algorithms to efficiently compute the causal effects, as well as scale the approach to data with large dimensionality. We also show how this method can be used for recurrent neural networks. We report experimental results on both simulated and real datasets showcasing the promise and usefulness of the proposed algorithm.