Learning a policy using only observational data is challenging because the distribution of states it induces at execution time may differ from the distribution observed during training. We propose to train a policy by unrolling a learned model of the environment dynamics over multiple time steps while explicitly penalizing two costs: the original cost the policy seeks to optimize, and an uncertainty cost which represents its divergence from the states it is trained on. We measure this second cost by using the uncertainty of the dynamics model about its own predictions, using recent ideas from uncertainty estimation for deep networks. We evaluate our approach using a large-scale observational dataset of driving behavior recorded from traffic cameras, and show that we are able to learn effective driving policies from purely observational data, with no environment interaction.

Imitation learning is a control design paradigm that seeks to learn a control policy reproducing demonstrations from experts. By substituting expert's demonstrations for optimal behaviours, the same paradigm leads to the design of control policies closely approximating the optimal state-feedback. This approach requires training a machine learning algorithm (in our case deep neural networks) directly on state-control pairs originating from optimal trajectories. We have shown in previous work that, when restricted to relatively low-dimensional state and control spaces, this approach is very successful in several deterministic, non-linear problems in continuous-time. In this work, we refine our previous studies using as test case a simple quadcopter model with quadratic and time-optimal objective functions. We describe in detail the best learning pipeline we have developed and that is able to approximate via deep neural networks the state-feedback map to a very high accuracy. We introduce the use of the softplus activation function in the hidden units showing how it results in a smoother control profile whilst retaining the benefits of ReLUs. We show how to evaluate the optimality of the trained state-feedback, and find that already with two layers the objective function reached and its optimal value differ by less than one percent. We later consider also an additional metric linked to the system asymptotic behaviour - time taken to converge to the policy's fixed point. With respect to these metrics, we show that improvements in the mean average error do not necessarily correspond to significant improvements.

Deep learning solutions are being increasingly used in mobile applications. Although there are many open-source software tools for the development of deep learning solutions, there are no guidelines in one place in a unified manner for using these tools towards real-time deployment of these solutions on smartphones. From the variety of available deep learning tools, the most suited ones are used in this paper to enable real-time deployment of deep learning inference networks on smartphones. A uniform flow of implementation is devised for both Android and iOS smartphones. The advantage of using multi-threading to achieve or improve real-time throughputs is also showcased. A benchmarking framework consisting of accuracy, CPU/GPU consumption and real-time throughput is considered for validation purposes. The developed deployment approach allows deep learning models to be turned into real-time smartphone apps with ease based on publicly available deep learning and smartphone software tools. This approach is applied to six popular or representative convolutional neural network models and the validation results based on the benchmarking metrics are reported.

We analyze the joint probability distribution on the lengths of the vectors of hidden variables in different layers of a fully connected deep network, when the weights and biases are chosen randomly according to Gaussian distributions, and the input is in $\{ -1, 1\}^N$. We show that, if the activation function $\phi$ satisfies a minimal set of assumptions, satisfied by all activation functions that we know that are used in practice, then, as the width of the network gets large, the `length process' converges in probability to a length map that is determined as a simple function of the variances of the random weights and biases, and the activation function $\phi$. We also show that this convergence may fail for $\phi$ that violate our assumptions.

In this note we discuss a common misconception, namely that embeddings are always used to reduce the dimensionality of the item space. We show that when we measure dimensionality in terms of information entropy then the embedding of sparse probability distributions, that can be used to represent sparse features or data, may or not reduce the dimensionality of the item space. However, the embeddings do provide a different and often more meaningful representation of the items for a particular task at hand. Also, we give upper bounds and more precise guidelines for choosing the embedding dimension.

Acquisition of Magnetic Resonance Imaging (MRI) scans can be accelerated by under-sampling in k-space (i.e., the Fourier domain). In this paper, we consider the problem of optimizing the sub-sampling pattern in a data-driven fashion. Since the reconstruction model's performance depends on the sub-sampling pattern, we combine the two problems. For a given sparsity constraint, our method optimizes the sub-sampling pattern and reconstruction model, using an end-to-end learning strategy. Our algorithm learns from full-resolution data that are under-sampled retrospectively, yielding a sub-sampling pattern and reconstruction model that are customized to the type of images represented in the training data. The proposed method, which we call LOUPE (Learning-based Optimization of the Under-sampling PattErn), was implemented by modifying a U-Net, a widely-used convolutional neural network architecture, that we append with the forward model that encodes the under-sampling process. Our experiments with T1-weighted structural brain MRI scans show that the optimized sub-sampling pattern can yield significantly more accurate reconstructions compared to standard random uniform, variable density or equispaced under-sampling schemes.

Stochastic computation graphs (SCGs) provide a formalism to represent structured optimization problems arising in artificial intelligence, including supervised, unsupervised, and reinforcement learning. Previous work has shown that an unbiased estimator of the gradient of the expected loss of SCGs can be derived from a single principle. However, this estimator often has high variance and requires a full model evaluation per data point, making this algorithm costly in large graphs. In this work, we address these problems by generalizing concepts from the reinforcement learning literature. We introduce the concepts of value functions, baselines and critics for arbitrary SCGs, and show how to use them to derive lower-variance gradient estimates from partial model evaluations, paving the way towards general and efficient credit assignment for gradient-based optimization. In doing so, we demonstrate how our results unify recent advances in the probabilistic inference and reinforcement learning literature.

Ultrasound (US) imaging is based on the time-reversal principle, in which individual channel RF measurements are back-propagated and accumulated to form an image after applying specific delays. While this time reversal is usually implemented as a delay-and-sum (DAS) beamformer, the image quality quickly degrades as the number of measurement channels decreases. To address this problem, various types of adaptive beamforming techniques have been proposed using predefined models of the signals. However, the performance of these adaptive beamforming approaches degrade when the underlying model is not sufficiently accurate. Here, we demonstrate for the first time that a single universal deep beamformer trained using a purely data-driven way can generate significantly improved images over widely varying aperture and channel subsampling patterns. In particular, we design an end-to-end deep learning framework that can directly process sub-sampled RF data acquired at different subsampling rate and detector configuration to generate high quality ultrasound images using a single beamformer. Experimental results using B-mode focused ultrasound confirm the efficacy of the proposed methods.

If you want to tell people the truth, make them laugh, otherwise they'll kill you. (source unclear) Machine learning and deep learning are the technologies of the day for developing intelligent automatic systems. However, a key hurdle for progress in the field is the literature itself: we often encounter papers that report results that are difficult to reconstruct or reproduce, results that mis-represent the performance of the system, or contain other biases that limit their validity. In this semi-humorous article, we discuss issues that arise in running and reporting results of machine learning experiments. The purpose of the article is to provide a list of watch out points for researchers to be aware of when developing machine learning models or writing and reviewing machine learning papers.

With deep neural networks providing state-of-the-art machine learning models for numerous machine learning tasks, quantifying the robustness of these models has become an important area of research. However, most of the research literature merely focuses on the \textit{worst-case} setting where the input of the neural network is perturbed with noises that are constrained within an $\ell_p$ ball; and several algorithms have been proposed to compute certified lower bounds of minimum adversarial distortion based on such worst-case analysis. In this paper, we address these limitations and extend the approach to a \textit{probabilistic} setting where the additive noises can follow a given distributional characterization. We propose a novel probabilistic framework PROVEN to PRObabilistically VErify Neural networks with statistical guarantees -- i.e., PROVEN certifies the probability that the classifier's top-1 prediction cannot be altered under any constrained $\ell_p$ norm perturbation to a given input. Importantly, we show that it is possible to derive closed-form probabilistic certificates based on current state-of-the-art neural network robustness verification frameworks. Hence, the probabilistic certificates provided by PROVEN come naturally and with almost no overhead when obtaining the worst-case certified lower bounds from existing methods such as Fast-Lin, CROWN and CNN-Cert. Experiments on small and large MNIST and CIFAR neural network models demonstrate our probabilistic approach can achieve up to around $75\%$ improvement in the robustness certification with at least a $99.99\%$ confidence compared with the worst-case robustness certificate delivered by CROWN.