Abstract

Variational autoencoders were proven successful in domains such as computer vision and speech processing. Their adoption for modeling user preferences is still unexplored, although recently it is starting to gain attention in the current literature. In this work, we propose a model which extends variational autoencoders by exploiting the rich information present in the past preference history. We introduce a recurrent version of the VAE, where instead of passing a subset of the whole history regardless of temporal dependencies, we rather pass the consumption sequence subset through a recurrent neural network. At each time-step of the RNN, the sequence is fed through a series of fully-connected layers, the output of which models the probability distribution of the most likely future preferences. We show that handling temporal information is crucial for improving the accuracy of the VAE: In fact, our model beats the current state-of-the-art by valuable margins because of its ability to capture temporal dependencies among the user-consumption sequence using the recurrent encoder still keeping the fundamentals of variational autoencoders intact.

Abstract

Encoding models can reveal and decode neural representations in the visual and semantic domains. However, a thorough understanding of how distributed information in auditory cortices and temporal evolution of music contribute to model performance is still lacking in the musical domain. We measured fMRI responses during naturalistic music listening and constructed a two-stage approach that frst mapped musical features in auditory cortices and then decoded novel musical pieces. We then probed the infuence of stimuli duration (number of time points) and spatial extent (number of voxels) on decoding accuracy. Our approach revealed a linear increase in accuracy with duration and a point of optimal model performance for the spatial extent. We further showed that Shannon entropy is a driving factor, boosting accuracy up to 95% for music with highest information content. These fndings provide key insights for future decoding and reconstruction algorithms and open new venues for possible clinical applications.

Abstract

There has been growing interest towards naturalistic neuroimaging experiments, which deepen our understanding of how human brain processes and integrates incoming streams of multifaceted sensory information, as commonly occurs in real world. Music is a good example of such complex continuous phenomenon. In a few recent fMRI studies examining neural correlates of music in continuous listening settings, multiple perceptual attributes of music stimulus were represented by a set of high-level features, produced as the linear combination of the acoustic descriptors computationally extracted from the stimulus audio. New method: fMRI data from naturalistic music listening experiment were employed here. Kernel principal component analysis (KPCA) was applied to acoustic descriptors extracted from the stimulus audio to generate a set of nonlinear stimulus features. Subsequently, perceptual and neural correlates of the generated high-level features were examined.

Abstract

A challenging problem in cognitive neuroscience is to relate the structural connectivity (SC) to the functional connectivity (FC) to better understand how large-scale network dynamics underlying human cognition emerges from the relatively fxed SC architecture. Recent modeling attempts point to the possibility of a single difusion kernel giving a good estimate of the FC. We highlight the shortcomings of the single-difusion-kernel model (SDK) and propose a multi-scale difusion scheme. Our multiscale model is formulated as a reaction-difusion system giving rise to spatio-temporal patterns on a fxed topology. We hypothesize the presence of inter-regional co-activations (latent parameters) that combine difusion kernels at multiple scales to characterize how FC could arise from SC. We formulated a multiple kernel learning (MKL) scheme to estimate the latent parameters from training data. Our model is analytically tractable and complex enough to capture the details of the underlying biological phenomena. The parameters learned by the MKL model lead to highly accurate predictions of subject-specifc FCs from test datasets at a rate of 71%, surpassing the performance of the existing linear and non-linear models. We provide an example of how these latent parameters could be used to characterize age-specifc reorganization in the brain structure and function.

Abstract

Extreme learning machine (ELM) is a new algorithm for training single-hidden layer feedforward neural networks which provides good performance as well as fast learning speed. ELM tends to produce good generalization performance with large number of hidden neurons as the input weights and hidden neurons biases are randomly initialized and remain unchanged during the learning process, and the output weights are analytically determined. In this paper, two swarm intelligence based metaheuristic techniques, viz. Artificial Bee Colony (ABC) and Invasive Weed Optimization (IWO) are proposed for tuning the input weights and hidden biases. The proposed approaches are called ABC-ELM and IWO-ELM in which the input weights and hidden biases are selected using ABC and IWO respectively and the output weights are computed using the Moore-Penrose (MP) generalized inverse. The proposed approaches are tested on different benchmark classification data sets and simulations show that the proposed approaches obtain good generalization performance in comparison to the other techniques available in the literature.

Abstract

Brain connectivity analysis has provided crucial insights to pinpoint the differences between autistic and typically developing (TD) children during development. The aim of this study is to investigate the functional connectomics of autism spectrum disorder (ASD) versus TD and underpin the effects of development, disease, and their interactions on the observed atypical brain connectivity patterns. Resting-state functional magnetic resonance imaging (rs-fMRI) from the Autism Brain Imaging Data Exchange (ABIDE) data set, which is stratified into two cohorts: children (9–12 years) and adolescents (13–16 years), is used for the analysis. Differences in various graph theoretical network measures are calculated between ASD and TD in each group. Furthermore, two-factor analysis of variance test is used to study the effect of age, disease, and their interaction on the network measures and the network edges. Furthermore, the differences in connection strength between TD and ASD subjects are assessed using network-based statistics. The results showed that ASD exhibits increased functional integration at the expense of decreased functional segregation. In ASD adolescents, there is a significant decrease in odularity suggesting a less robust modular organization, and an increase in participation coefficient suggesting more random integration and widely distributed connection edges. Furthermore, there is significant hypoconnectivity observed in the adolescent group especially in the default mode network, while the children group shows both hyper- and hypoconnectivity. This study lends support to a model of global atypical connections and further identifies functional networks and areas that are independently affected by age, disease, and their interaction.

Abstract

Encoding models of functional magnetic resonance imaging (fMRI) data attempt to learn a forward mapping that relates stimuli to the corresponding brain activation. Computational tractability usually forces current encoding as well as decoding solutions to typically consider only a small subset of voxels from the actual 3D volume of activation. Further, while brain decoding has received wider attention, there have been only a few attempts at constructing encoding solutions in the extant neuroimaging literature. In this paper, we present a deep autoencoder consisting of convolutional neural networks in tandem with long short-term memory (CNNLSTM) model. The model is trained on fMRI slice sequences and predicts the entire brain volume rather than a small subset of voxels from the information in stimuli (text and image). We argue that the resulting solution avoids the problem of devising encoding models based on a rule-based selection of informative voxels and the concomitant issue of wide spatial variability of such voxels across participants. The perturbation experiments indicate that the proposed deep encoder indeed learns to predict brain activations with high spatial accuracy. On the challenging universal decoder imaging datasets (Pereira et al., 2018), our model yielded encouraging results.

Abstract

Previous research shows that performance is better when a high pitch is responded with up or right responses and a low pitch is responded with down or left responses, called the spatial-pitch association of response codes (SPARC) effect. Despite the intuitive coupling of perception-action, studies investigating the SPARC effect have, however, used feedback to manipulate the stimulus-response mapping. Feedback contradicts the purpose of intuitive stimulus-response mapping by enabling short-term learning. This study primarily investigates the role of feedback on SPARC effect. We believe that feedback can facilitate incongruent mapping and can, therefore, reduce the cost between incongruent and congruent mapping resulting in a diminished SPARC effect. Our results, however, show that feedback has no influence on the SPARC effect indicating that long-term associations can not be overcome by short-term learning due to robust perception-action coupling. Further, unlike previous studies, we observed a strong horizontal SPARC effect in nonmusicians as well.

Abstract

POS Tagging serves as a preliminary task for many NLP applications. Kannada is a relatively poor Indian language with very limited number of quality NLP tools available for use. An accurate and reliable POS Tagger is essential for many NLP tasks like shallow parsing, dependency parsing, sentiment analysis, named entity recognition. We present a statistical POS tagger for Kannada using different machine learning and neural network models. Our Kannada POS tagger outperforms the state-of-the-art Kannada POS tagger by 6%. Our contribution in this paper is three folds - building a generic POS Tagger, comparing the performances of different modeling techniques, exploring the use of character and word embeddings together for Kannada POS Tagging.

Abstract

The Part-Of-Speech tag of a word can provide crucial information for a large number of tasks, and so, it is of utmost importance that the POS tagged data is accurate. However, manually checking the data is a tedious and time consuming task. Thus, there is a need for an Automatic Error Correction and Validation model for any POS Tagged Data. In this paper,we work towards achieving the aforementioned goal for Hindi POS Tagging. This is achieved by using an ensemble model consisting of three POS Tagging Models. Based on the predictions made by the three models, and the POS tag present in the dataset, the ensemble model predicts the presence of an error. The POS tagging models explored were the Hidden Markov Model, Support Vector Machine, Conditional Random Fields, Long Short Term Memory (LSTM) Networks, Bidirectional LSTM Networks, and Logistic Regression. A Fully Connected Neural Network was used to build the ensemble model, and it achieved an accuracy of 94.02%.