Deep feedforward and recurrent networks have achieved impressive results in many perception and language processing applications. This success is partially attributed to architectural innovations such as convolutional and long short-term memory networks. The main motivation for these architectural innovations is that they capture better domain knowledge, and importantly are easier to optimize than more basic architectures. Recently, more complex architectures such as Neural Turing Machines and Memory Networks have been proposed for tasks including question answering and general computation, creating a new set of optimization challenges. In this paper, we discuss a low-overhead and easy-to-implement technique of adding gradient noise which we find to be surprisingly effective when training these very deep architectures. The technique not only helps to avoid overfitting, but also can result in lower training loss. This method alone allows a fully-connected 20-layer deep network to be trained with standard gradient descent, even starting from a poor initialization. We see consistent improvements for many complex models, including a 72% relative reduction in error rate over a carefully-tuned baseline on a challenging question-answering task, and a doubling of the number of accurate binary multiplication models learned across 7,000 random restarts. We encourage further application of this technique to additional complex modern architectures.

In some speaker recognition scenarios we find conversations recorded simultaneously over multiple channels. That is the case of the interviews in the NIST SRE dataset. To take advantage of that, we propose a modification of the PLDA model that considers two different inter-session variability terms. The first term is tied between all the recordings belonging to the same conversation whereas the second is not. Thus, the former mainly intends to capture the variability due to the phonetic content of the conversation while the latter tries to capture the channel variability. In this document, we derive the equations for this model. This model was applied in the paper "Handling Recordings Acquired Simultaneously over Multiple Channels with PLDA" published at Interspeech 2013.

Normalized nonnegative models assign probability distributions to users and random variables to items; see [Stark, 2015]. Rating an item is regarded as sampling the random variable assigned to the item with respect to the distribution assigned to the user who rates the item. Models of that kind are highly expressive. For instance, using normalized nonnegative models we can understand users' preferences as mixtures of interpretable user stereotypes, and we can arrange properties of users and items in a hierarchical manner. These features would not be useful if the predictive power of normalized nonnegative models was poor. Thus, we analyze here the performance of normalized nonnegative models for top-N recommendation and observe that their performance matches the performance of methods like PureSVD which was introduced in [Cremonesi et al., 2010]. We conclude that normalized nonnegative models not only provide accurate recommendations but they also deliver (for free) representations that are interpretable. We deepen the discussion of normalized nonnegative models by providing further theoretical insights. In particular, we introduce total variational distance as an operational similarity measure, we discover scenarios where normalized nonnegative models yield unique representations of users and items, we prove that the inference of optimal normalized nonnegative models is NP-hard and finally, we discuss the relationship between normalized nonnegative models and nonnegative matrix factorization.

We introduce the "NoBackTrack" algorithm to train the parameters of dynamical systems such as recurrent neural networks. This algorithm works in an online, memoryless setting, thus requiring no backpropagation through time, and is scalable, avoiding the large computational and memory cost of maintaining the full gradient of the current state with respect to the parameters. The algorithm essentially maintains, at each time, a single search direction in parameter space. The evolution of this search direction is partly stochastic and is constructed in such a way to provide, at every time, an unbiased random estimate of the gradient of the loss function with respect to the parameters. Because the gradient estimate is unbiased, on average over time the parameter is updated as it should. The resulting gradient estimate can then be fed to a lightweight Kalman-like filter to yield an improved algorithm. For recurrent neural networks, the resulting algorithms scale linearly with the number of parameters. Small-scale experiments confirm the suitability of the approach, showing that the stochastic approximation of the gradient introduced in the algorithm is not detrimental to learning. In particular, the Kalman-like version of NoBackTrack is superior to backpropagation through time (BPTT) when the time span of dependencies in the data is longer than the truncation span for BPTT.

We introduce Embed to Control (E2C), a method for model learning and control of non-linear dynamical systems from raw pixel images. E2C consists of a deep generative model, belonging to the family of variational autoencoders, that learns to generate image trajectories from a latent space in which the dynamics is constrained to be locally linear. Our model is derived directly from an optimal control formulation in latent space, supports long-term prediction of image sequences and exhibits strong performance on a variety of complex control problems.

State-of-the-art speaker recognition relays on models that need a large amount of training data. This models are successful in tasks like NIST SRE because there is sufficient data available. However, in real applications, we usually do not have so much data and, in many cases, the speaker labels are unknown. We present a method to adapt a PLDA model from a domain with a large amount of labeled data to another with unlabeled data. We describe a generative model that produces both sets of data where the unknown labels are modeled like latent variables. We used variational Bayes to estimate the hidden variables. Here, we derive the equations for this model. This model has been used in the papers: "UNSUPERVISED ADAPTATION OF PLDA BY USING VARIATIONAL BAYES METHODS" publised at ICASSP 2014, "Unsupervised Training of PLDA with Variational Bayes" published at Iberspeech 2014, and "VARIATIONAL BAYESIAN PLDA FOR SPEAKER DIARIZATION IN THE MGB CHALLENGE" published at ASRU 2015.

Although the traits emerged in a mass gathering are often non-deliberative, the act of mass impulse may lead to irre- vocable crowd disasters. The two-fold increase of carnage in crowd since the past two decades has spurred significant advances in the field of computer vision, towards effective and proactive crowd surveillance. Computer vision stud- ies related to crowd are observed to resonate with the understanding of the emergent behavior in physics (complex systems) and biology (animal swarm). These studies, which are inspired by biology and physics, share surprisingly common insights, and interesting contradictions. However, this aspect of discussion has not been fully explored. Therefore, this survey provides the readers with a review of the state-of-the-art methods in crowd behavior analysis from the physics and biologically inspired perspectives. We provide insights and comprehensive discussions for a broader understanding of the underlying prospect of blending physics and biology studies in computer vision.

Deep Convolutional Neural Networks (DCNN) has shown excellent performance in a variety of machine learning tasks. This manuscript presents Deep Convolutional Neural Fields (DeepCNF), a combination of DCNN with Conditional Random Field (CRF), for sequence labeling with highly imbalanced label distribution. The widely-used training methods, such as maximum-likelihood and maximum labelwise accuracy, do not work well on highly imbalanced data. To handle this, we present a new training algorithm called maximum-AUC for DeepCNF. That is, we train DeepCNF by directly maximizing the empirical Area Under the ROC Curve (AUC), which is an unbiased measurement for imbalanced data. To fulfill this, we formulate AUC in a pairwise ranking framework, approximate it by a polynomial function and then apply a gradient-based procedure to optimize it. We then test our AUC-maximized DeepCNF on three very different protein sequence labeling tasks: solvent accessibility prediction, 8-state secondary structure prediction, and disorder prediction. Our experimental results confirm that maximum-AUC greatly outperforms the other two training methods on 8-state secondary structure prediction and disorder prediction since their label distributions are highly imbalanced and also have similar performance as the other two training methods on the solvent accessibility prediction problem which has three equally-distributed labels. Furthermore, our experimental results also show that our AUC-trained DeepCNF models greatly outperform existing popular predictors of these three tasks.

Diffusion Maps framework is a kernel based method for manifold learning and data analysis that defines diffusion similarities by imposing a Markovian process on the given dataset. Analysis by this process uncovers the intrinsic geometric structures in the data. Recently, it was suggested to replace the standard kernel by a measure-based kernel that incorporates information about the density of the data. Thus, the manifold assumption is replaced by a more general measure-based assumption. The measure-based diffusion kernel incorporates two separate independent representations. The first determines a measure that correlates with a density that represents normal behaviors and patterns in the data. The second consists of the analyzed multidimensional data points. In this paper, we present a representation framework for data analysis of datasets that is based on a closed-form decomposition of the measure-based kernel. The proposed representation preserves pairwise diffusion distances that does not depend on the data size while being invariant to scale. For a stationary data, no out-of-sample extension is needed for embedding newly arrived data points in the representation space. Several aspects of the presented methodology are demonstrated on analytically generated data.

In clinical and neuroscientific studies, systematic differences between two populations of brain networks are investigated in order to characterize mental diseases or processes. Those networks are usually represented as graphs built from neuroimaging data and studied by means of graph analysis methods. The typical machine learning approach to study these brain graphs creates a classifier and tests its ability to discriminate the two populations. In contrast to this approach, in this work we propose to directly test whether two populations of graphs are different or not, by using the kernel two-sample test (KTST), without creating the intermediate classifier. We claim that, in general, the two approaches provides similar results and that the KTST requires much less computation. Additionally, in the regime of low sample size, we claim that the KTST has lower frequency of Type II error than the classification approach. Besides providing algorithmic considerations to support these claims, we show strong evidence through experiments and one simulation.