No abstract

Unsupervised feature extractors are known to perform an efficient and discriminative representation of data. Insight into the mappings they perform and human ability to understand them, however, remain very limited. This is especially prominent when multilayer deep learning architectures are used. This paper demonstrates how to remove these bottlenecks within the architecture of non-negativity constrained autoencoder. It is shown that using both L1 and L2 regularizations that induce non-negativity of weights, most of the weights in the network become constrained to be non-negative, thereby resulting into a more understandable structure with minute deterioration in classification accuracy. Also, this proposed approach extracts features that are more sparse and produces additional output layer sparsification. The method is analyzed for accuracy and feature interpretation on the MNIST data, the NORB normalized uniform object data, and the Reuters text categorization data set.

Autism spectrum disorder (ASD) is a neuro-developmental disorder associated with social impairments, communication difficulties, and restricted and repetitive behaviors. Yet, there is no confirmed cause identified for ASD. Studying the functional connectivity of the brain is an emerging technique used in diagnosing and understanding ASD. In this study, we obtained the resting state functional MRI data of 283 subjects from the National Database of Autism Research (NDAR). An automated autism diagnosis system was built using the data from NDAR. The proposed system is machine learning based. Power spectral densities (PSDs) of time courses corresponding to the spatial activation areas are used as input features, feeds them to a stacked autoencoder then builds a classifier using probabilistic support vector machines. Over the used dataset, around 90% of sensitivity, specificity and accuracy was achieved by our machine learning system. Moreover, the system generalization ability was checked over two different prevalence values, one for the general population and the other for the of high risk population, and the system proved to be very generalizable, especially among the population of high risk. The proposed system generates a full personalized report for each subject, along with identifying the global differences between ASD and typically developed (TD) subjects and its ability to diagnose autism. It shows the impacted areas and the severity of implications. From the clinical aspect, this report is considered very valuable as it helps in both predicting and understanding behavior of autistic subjects. Moreover, it helps in designing a plan for personalized treatment per each individual subject. The proposed work is taking a step towards achieving personalized medicine in autism which is the ultimate goal of our group’s research efforts in this area.

No abstract

Random deployment of sensor nodes is susceptible to initial communication hole, even when the network is densely populated. However, eliminating holes using structural deployment poses its difficulties. In either case, the resulting coverage holes can degrade overall network performance and lifetime. Many solutions utilizing Relay Nodes (RNs) have been proposed to alleviate this problem. In this regard, one of the recent solutions proposed using Artificial Bee Colony (ABC) to deploy RNs. This paper proposes RN deployment using two other evolutionary techniques -Gravitational Search Algorithm (GSA) and Differential Evolution (DE) and compares them with existing solution that uses ABC. These popular optimization tools are deployed to optimize the positions of relay nodes for lifetime maximization. Proposed algorithms guarantee satisfactory RNs utilization while maintaining desired connectivity level. It is shown that DE-based deployment improves the network lifetime better than other optimization heuristics considered.