Abstract--We present an interpretable neural network for predicting an important clinical outcome (1-year mortality) from multi-modal Electronic Health Record (EHR) data. Our approach buildson prior multi-modal machine learning models by now enabling visualization of how individual factors contribute to the overall outcome risk, assuming other factors remain constant, which was previously impossible. We demonstrate the value of this approach using a large multi-modal clinical dataset including both EHR data and 31,278 echocardiographic videos of the heart from 26,793 patients. We generated separate models for (i) clinical data only (CD) (e.g. The interpretable multi-modal model maintained performance compared to non-interpretable models (Random Forest, XG-Boost), and also performed significantly better than a model using a single modality (average AUC 0.82). Clinically relevant insights and multi-modal variable importance rankings were also facilitated by the new model, which have previously been impossible. I. INTRODUCTION The adoption of Electronic Health Records (EHR) in medicine has facilitated the collection of massive amounts of clinical data which can be used to develop highly accurate risk models that physicians can use to guide medical decision making. To take full advantage of the available EHR data, these models, similar to a physician, need to be able to handle multiple modalities as inputs. For example, both tabular data such as laboratory measurements and pixel data from clinical images should be readily incorporated. This basic framework is shown in Figure 1. As documented in [1], [2], [3], [4], precision medicine can benefit greatly from development of these risk models. The proliferation of these models has prompted scrutiny from the medical community, which demands clinical validity and interpretability to improve usefulness [5] and inclusion of all relevant predictors (or, conversely, explanation when a relevant data input is excluded) [6]. Thus, any medical risk model should be interpretable and facilitate understanding of the various contributions of different inputs towards the overall risk assessment.

Convolutional Neural Networks (CNNs) are widely used to solve classification tasks in computer vision. However, they can be tricked into misclassifying specially crafted `adversarial' samples -- and samples built to trick one model often work alarmingly well against other models trained on the same task. In this paper we introduce Sitatapatra, a system designed to block the transfer of adversarial samples. It diversifies neural networks using a key, as in cryptography, and provides a mechanism for detecting attacks. What's more, when adversarial samples are detected they can typically be traced back to the individual device that was used to develop them. The run-time overheads are minimal permitting the use of Sitatapatra on constrained systems.

Fitting machine learning models in the low-data limit is challenging. The main challenge is to obtain suitable prior knowledge and encode it into the model, for instance in the form of a Gaussian process prior. Recent advances in meta-learning offer powerful methods for extracting such prior knowledge from data acquired in related tasks. When it comes to meta-learning in Gaussian process models, approaches in this setting have mostly focused on learning the kernel function of the prior, but not on learning its mean function. In this work, we propose to parameterize the mean function of a Gaussian process with a deep neural network and train it with a meta-learning procedure. We present analytical and empirical evidence that mean function learning can be superior to kernel learning alone, particularly if data is scarce.

Despite the recent popularity of deep generative state space models, few comparisons have been made between network architectures and the inference steps of the Bayesian filtering framework -- with most models simultaneously approximating both state transition and update steps with a single recurrent neural network (RNN). In this paper, we introduce the Recurrent Neural Filter (RNF), a novel recurrent variational autoencoder architecture that learns distinct representations for each Bayesian filtering step, captured by a series of encoders and decoders. Testing this on three real-world time series datasets, we demonstrate that decoupling representations not only improves the accuracy of one-step-ahead forecasts while providing realistic uncertainty estimates, but also facilitates multistep prediction through the separation of encoder stages.

While extraordinary progress has been made towards developing neural network architectures for classification tasks, commonly used loss functions such as the multi-category cross entropy loss are inadequate for ranking and ordinal regression problems. To address this issue, approaches have been developed that transform ordinal target variables series of binary classification tasks, resulting in robust ranking algorithms with good generalization performance. However, to model ordinal information appropriately, ideally, a rank-monotonic prediction function is required such that confidence scores are ordered and consistent. We propose a new framework (Consistent Rank Logits, CORAL) with theoretical guarantees for rank-monotonicity and consistent confidence scores. Through parameter sharing, our framework benefits from low training complexity and can easily be implemented to extend common convolutional neural network classifiers for ordinal regression tasks. Furthermore, our empirical results support the proposed theory and show a substantial improvement compared to the current state-of-the-art ordinal regression method for age prediction from face images.

Successful multimodal search and retrieval requires the automatic understanding of semantic cross-modal relations, which, however, is still an open research problem. Previous work has suggested the metrics cross-modal mutual information and semantic correlation to model and predict cross-modal semantic relations of image and text. In this paper, we present an approach to predict the (cross-modal) relative abstractness level of a given image-text pair, that is whether the image is an abstraction of the text or vice versa. For this purpose, we introduce a new metric that captures this specific relationship between image and text at the Abstractness Level (ABS). We present a deep learning approach to predict this metric, which relies on an autoencoder architecture that allows us to significantly reduce the required amount of labeled training data. A comprehensive set of publicly available scientific documents has been gathered. Experimental results on a challenging test set demonstrate the feasibility of the approach.

Rapid improvements in machine learning over the past decade are beginning to have far-reaching effects. For communications, engineers with limited domain expertise can now use off-the-shelf learning packages to design high-performance systems based on simulations. Prior to the current revolution in machine learning, the majority of communication engineers were quite aware that system parameters (such as filter coefficients) could be learned using stochastic gradient descent. It was not at all clear, however, that more complicated parts of the system architecture could be learned as well. In this paper, we discuss the application of machine-learning techniques to two communications problems and focus on what can be learned from the resulting systems. We were pleasantly surprised that the observed gains in one example have a simple explanation that only became clear in hindsight. In essence, deep learning discovered a simple and effective strategy that had not been considered earlier.

Reinforcement learning (RL) algorithms allow agents to learn skills and strategies to perform complex tasks without detailed instructions or expensive labelled training examples. That is, RL agents can learn, as we learn. Given the importance of learning in our intelligence, RL has been thought to be one of key components to general artificial intelligence, and recent breakthroughs in deep reinforcement learning suggest that neural networks (NN) are natural platforms for RL agents. However, despite the efficiency and versatility of NN-based RL agents, their decision-making remains incomprehensible, reducing their utilities. To deploy RL into a wider range of applications, it is imperative to develop explainable NN-based RL agents. Here, we propose a method to derive a secondary comprehensible agent from a NN-based RL agent, whose decision-makings are based on simple rules. Our empirical evaluation of this secondary agent's performance supports the possibility of building a comprehensible and transparent agent using a NN-based RL agent.

Bayesian online changepoint detection (BOCPD) (Adams & MacKay, 2007) offers a rigorous and viable way to identity changepoints in complex systems. In this work, we introduce a Stein variational online changepoint detection (SVOCD) method to provide a computationally tractable generalization of BOCPD beyond the exponential family of probability distributions. We integrate the recently developed Stein variational Newton (SVN) method (Detommaso et al., 2018) and BOCPD to offer a full online Bayesian treatment for a large number of situations with significant importance in practice. We apply the resulting method to two challenging and novel applications: Hawkes processes and long short-term memory (LSTM) neural networks. In both cases, we successfully demonstrate the efficacy of our method on real data.

We report an extension of a Keras Model, called CTCModel, to perform the Connectionist Temporal Classification (CTC) in a transparent way. Combined with Recurrent Neural Networks, the Connectionist Temporal Classification is the reference method for dealing with unsegmented input sequences, i.e. with data that are a couple of observation and label sequences where each label is related to a subset of observation frames. CTCModel makes use of the CTC implementation in the Tensorflow backend for training and predicting sequences of labels using Keras. It consists of three branches made of Keras models: one for training, computing the CTC loss function; one for predicting, providing sequences of labels; and one for evaluating that returns standard metrics for analyzing sequences of predictions.