We compare a variety of models for predicting early hospital readmissions. Performance of existing models is insufficient for practical applications. Random forests and deep neural networks perform best in terms of AUC. Models fit to homogeneous patient subgroups typically outperform global models. Risk sharing arrangements between hospitals and payers together with penalties imposed by the Centers for Medicare and Medicaid (CMS) are driving an interest in decreasing early readmissions.

We have recently shown that deep Long Short-Term Memory (LSTM) recurrent neural networks (RNNs) outperform feed forward deep neural networks (DNNs) as acoustic models for speech recognition. More recently, we have shown that the performance of sequence trained context dependent (CD) hidden Markov model (HMM) acoustic models using such LSTM RNNs can be equaled by sequence trained phone models initialized with connectionist temporal classification (CTC). In this paper, we present techniques that further improve performance of LSTM RNN acoustic models for large vocabulary speech recognition. We show that frame stacking and reduced frame rate lead to more accurate models and faster decoding. CD phone modeling leads to further improvements.

Learning and inferring features that generate sensory input is a task continuously performed by cortex. In recent years, novel algorithms and learning rules have been proposed that allow neural network models to learn such features from natural images, written text, audio signals, etc. These networks usually involve deep architectures with many layers of hidden neurons. Here we review recent advancements in this area emphasizing, amongst other things, the processing of dynamical inputs by networks with hidden nodes and the role of single neuron models. These points and the questions they arise can provide conceptual advancements in understanding of learning in the cortex and the relationship between machine learning approaches to learning with hidden nodes and those in cortical circuits. Keywords: statistical models, deep learning, dynamics 1. Introduction Learning the fundamental features that generate sensory signals and having the ability to infer these features once given the sensory input is a crucial aspect of information processing. It would not be far fetched to hypothesize that the organization of cortical circuitry is largely evolved for performing such tasks, or to put it slightly differently, that our evolutionary history has stored memories of such features in the connections that form a large part of our brains. But how can a neuronal network learn and extract features that cause the experiences of our sensory organs, in a supervised or unsupervised manner?

Invariance to geometric transformations is a highly desirable property of automatic classifiers in many image recognition tasks. Nevertheless, it is unclear to which extent state-of-the-art classifiers are invariant to basic transformations such as rotations and translations. This is mainly due to the lack of general methods that properly measure such an invariance. In this paper, we propose a rigorous and systematic approach for quantifying the invariance to geometric transformations of any classifier. Our key idea is to cast the problem of assessing a classifier's invariance as the computation of geodesics along the manifold of transformed images. We propose the Manitest method, built on the efficient Fast Marching algorithm to compute the invariance of classifiers. Our new method quantifies in particular the importance of data augmentation for learning invariance from data, and the increased invariance of convolutional neural networks with depth. We foresee that the proposed generic tool for measuring invariance to a large class of geometric transformations and arbitrary classifiers will have many applications for evaluating and comparing classifiers based on their invariance, and help improving the invariance of existing classifiers.

Marginalising out uncertain quantities within the internal representations or parameters of neural networks is of central importance for a wide range of learning techniques, such as empirical, variational or full Bayesian methods. We set out to generalise fast dropout (Wang & Manning, 2013) to cover a wider variety of noise processes in neural networks. This leads to an efficient calculation of the marginal likelihood and predictive distribution which evades sampling and the consequential increase in training time due to highly variant gradient estimates. This allows us to approximate variational Bayes for the parameters of feed-forward neural networks. Inspired by the minimum description length principle, we also propose and experimentally verify the direct optimisation of the regularised predictive distribution. The methods yield results competitive with previous neural network based approaches and Gaussian processes on a wide range of regression tasks.

The fully connected layers of a deep convolutional neural network typically contain over 90% of the network parameters, and consume the majority of the memory required to store the network parameters. Reducing the number of parameters while preserving essentially the same predictive performance is critically important for operating deep neural networks in memory constrained environments such as GPUs or embedded devices. In this paper we show how kernel methods, in particular a single Fastfood layer, can be used to replace all fully connected layers in a deep convolutional neural network. This novel Fastfood layer is also end-to-end trainable in conjunction with convolutional layers, allowing us to combine them into a new architecture, named deep fried convolutional networks, which substantially reduces the memory footprint of convolutional networks trained on MNIST and ImageNet with no drop in predictive performance.

This work analyzes centered binary Restricted Boltzmann Machines (RBMs) and binary Deep Boltzmann Machines (DBMs), where centering is done by subtracting offset values from visible and hidden variables. We show analytically that (i) centering results in a different but equivalent parameterization for artificial neural networks in general, (ii) the expected performance of centered binary RBMs/DBMs is invariant under simultaneous flip of data and offsets, for any offset value in the range of zero to one, (iii) centering can be reformulated as a different update rule for normal binary RBMs/DBMs, and (iv) using the enhanced gradient is equivalent to setting the offset values to the average over model and data mean. Furthermore, numerical simulations suggest that (i) optimal generative performance is achieved by subtracting mean values from visible as well as hidden variables, (ii) centered RBMs/DBMs reach significantly higher log-likelihood values than normal binary RBMs/DBMs, (iii) centering variants whose offsets depend on the model mean, like the enhanced gradient, suffer from severe divergence problems, (iv) learning is stabilized if an exponentially moving average over the batch means is used for the offset values instead of the current batch mean, which also prevents the enhanced gradient from diverging, (v) centered RBMs/DBMs reach higher LL values than normal RBMs/DBMs while having a smaller norm of the weight matrix, (vi) centering leads to an update direction that is closer to the natural gradient and that the natural gradient is extremly efficient for training RBMs, (vii) centering dispense the need for greedy layer-wise pre-training of DBMs, (viii) furthermore we show that pre-training often even worsen the results independently whether centering is used or not, and (ix) centering is also beneficial for auto encoders.

Distributed word representations have a rising interest in NLP community. Most of existing models assume only one vector for each individual word, which ignores polysemy and thus degrades their effectiveness for downstream tasks. To address this problem, some recent work adopts multi-prototype models to learn multiple embeddings per word type. In this paper, we distinguish the different senses of each word by their latent topics. We present a general architecture to learn the word and topic embeddings efficiently, which is an extension to the Skip-Gram model and can model the interaction between words and topics simultaneously. The experiments on the word similarity and text classification tasks show our model outperforms state-of-the-art methods.

Distributed word representations have a rising interest in NLP community. Most of existing models assume only one vector for each individual word, which ignores polysemy and thus degrades their effectiveness for downstream tasks. To address this problem, some recent work adopts multi-prototype models to learn multiple embeddings per word type. In this paper, we distinguish the different senses of each word by their latent topics. We present a general architecture to learn the word and topic embeddings efficiently, which is an extension to the Skip-Gram model and can model the interaction between words and topics simultaneously. The experiments on the word similarity and text classification tasks show our model outperforms state-of-the-art methods.

The ability to accurately model a sentence at varying stages (e.g., word-phrase-sentence) plays a central role in natural language processing. As an effort towards this goal we propose a self-adaptive hierarchical sentence model (AdaSent). AdaSent effectively forms a hierarchy of representations from words to phrases and then to sentences through recursive gated local composition of adjacent segments. We design a competitive mechanism (through gating networks) to allow the representations of the same sentence to be engaged in a particular learning task (e.g., classification), therefore effectively mitigating the gradient vanishing problem persistent in other recursive models. Both qualitative and quantitative analysis shows that AdaSent can automatically form and select the representations suitable for the task at hand during training, yielding superior classification performance over competitor models on 5 benchmark data sets.