Matrix completion with rating data and auxiliary information for users and items is a challenging task in recommendation systems. In this paper, we propose an end-to-end architecture named Attention-based Graph Convolutional Network (AGCN) to embed both rating data and auxiliary information in a unified space, and subsequently learn low-rank dense representations via graph convolutional networks and attention layers. Compared to previous work, AGCN reduces computational complexity with Chebyshev polynomial graph filters. The introduced attention layer, which encourages weighing the neighbor information to learn more expressive structural graph representations, can improve the prediction accuracy, and lead to faster and more stable convergence. Experimental results show that our model can perform better and converge faster than current state-of-the-art methods on the real-world MovieLens and Flixster datasets.

The quest for biologically plausible deep learning is driven, not just by the desire to explain experimentally-observed properties of biological neural networks, but also by the hope of discovering more efficient methods for training artificial networks. In this paper, we propose a new algorithm named Variational Probably Flow (VPF), an extension of minimum probability flow for training binary Deep Boltzmann Machines (DBMs). We show that weight updates in VPF are local, depending only on the states and firing rates of the adjacent neurons. Unlike contrastive divergence, there is no need for Gibbs confabulations; and unlike backpropagation, alternating feedforward and feedback phases are not required. Moreover, the learning algorithm is effective for training DBMs with intra-layer connections between the hidden nodes. Experiments with MNIST and Fashion MNIST demonstrate that VPF learns reasonable features quickly, reconstructs corrupted images more accurately, and generates samples with a high estimated log-likelihood. Lastly, we note that, interestingly, if an asymmetric version of VPF exists, the weight updates directly explain experimental results in Spike-Timing-Dependent Plasticity (STDP).

Heterogeneous sensor data fusion is a challenging field that has gathered significant interest in recent years. In this paper, we propose a neural network-based multimodal data fusion framework named deep multimodal encoder (DME). Through our new objective function, both the intra- and inter-modal correlations of multimodal sensor data can be better exploited for recovering the missing values, and the shared representation learned can be used directly for prediction tasks. In experiments with real-world sensor data, DME shows remarkable ability for missing data imputation and new modality prediction. Compared with traditional algorithms such as kNN and Sparse-PCA, DME is more expressive, robust, and scalable to large datasets.

Compressive sensing has been successfully used for optimized operations in wireless sensor networks. However, raw data collected by sensors may be neither originally sparse nor easily transformed into a sparse data representation. This paper addresses the problem of transforming source data collected by sensor nodes into a sparse representation with a few nonzero elements. Our contributions that address three major issues include: 1) an effective method that extracts population sparsity of the data, 2) a sparsity ratio guarantee scheme, and 3) a customized learning algorithm of the sparsifying dictionary. We introduce an unsupervised neural network to extract an intrinsic sparse coding of the data. The sparse codes are generated at the activation of the hidden layer using a sparsity nomination constraint and a shrinking mechanism. Our analysis using real data samples shows that the proposed method outperforms conventional sparsity-inducing methods.

We present a data compression and dimensionality reduction scheme for data fusion and aggregation applications to prevent data congestion and reduce energy consumption at network connecting points such as cluster heads and gateways. Our in-network approach can be easily tuned to analyze the data temporal or spatial correlation using an unsupervised neural network scheme, namely the autoencoders. In particular, our algorithm extracts intrinsic data features from previously collected historical samples to transform the raw data into a low dimensional representation. Moreover, the proposed framework provides an error bound guarantee mechanism. We evaluate the proposed solution using real-world data sets and compare it with traditional methods for temporal and spatial data compression. The experimental validation reveals that our approach outperforms several existing wireless sensor network's data compression methods in terms of compression efficiency and signal reconstruction.

Sensor networks are increasingly being used to provide timely information about the physical, urban and human environment. Algorithms that depend on sensor data often assume that the readings are complete. However, node failures or communication breakdowns result in missing data entries, preventing the use of such algorithms. To impute these missing values, we propose a method of exploiting spatial correlations which is based on the sparse autoencoder and inspired by the conditional Restricted Boltzmann Machine that contested for the Netflix Prize. We modify the autoencoder to cope with missing data, and test it on data from a sensor testbed in Santander, Spain. We show that our algorithm extracts features from datasets with high proportions of missing data and uses these features to accurately and efficiently impute missing entries.

This paper presents a solution to the problem of monitoring a region of interest (RoI) using a set of nodes that is not sufficient to achieve the required degree of monitoring coverage. In particular, sensing coverage of wireless sensor networks (WSNs) is a crucial issue in projects due to failure of sensors. This scenario of limited funding hinders the traditional method of using mobile robots to move around the RoI to collect readings. Instead, our solution employs supervised neural networks to produce the values of the uncovered locations by extracting the non-linear relation among randomly deployed sensor nodes throughout the area. Moreover, we apply a hybrid backpropagation method to accelerate the learning convergence speed to a local minimum solution. We use a real-world data set from meteorological deployment for experimental validation and analysis.

We find the capacity pre-log of a temporally correlated Rayleigh block-fading single-input multiple-output (SIMO) channel in the noncoherent setting. It is well known that for block-lengthLand rank of the channel covariance matrix equal toQ, the capacity pre-log in the single-input single-output (SISO) case is given by 1-Q/L. Here,Q/Lcan be interpreted as the pre-log penalty incurred by channel uncertainty. Our main result reveals that, by adding only one receive antenna, this penalty can be reduced to 1/Land can, hence, be made to vanish for the block-lengthL→∞, even ifQ/Lremains constant asL→∞. Intuitively, even though the SISO channels between the transmit antenna and the two receive antennas are statistically independent, the transmit signal induces enough statistical dependence between the corresponding receive signals for the second receive antenna to be able to resolve the uncertainty associated with the first receive antenna's channel and thereby make the overall system appear coherent. The proof of our main theorem is based on a deep result from algebraic geometry known as Hironaka's Theorem on the Resolution of Singularities.

We establish a lower bound on the noncoherent capacity pre-log of a temporally correlated Rayleigh block-fading single-input multiple-output (SIMO) channel. Our result holds for arbitrary rank Q of the channel correlation matrix, arbitrary block-length L >; Q, and arbitrary number of receive antennas R, and includes the result in Morgenshtern et al. (2010) as a special case. It is well known that the capacity pre-log for this channel in the single-input single-output (SISO) case is given by 1-Q/L, where Q/L is the penalty incurred by channel uncertainty. Our result reveals that this penalty can be reduced to 1/L by adding only one receive antenna, provided that L ≥ 2Q - 1 and the channel correlation matrix satisfies mild technical conditions. The main technical tool used to prove our result is Hironaka's celebrated theorem on resolution of singularities in algebraic geometry.

In recent years, the research on multiple-input multiple-output (MIMO) broadcast channels has attracted much interest, especially since the discovery of the broadcast channel capacity achievable through the use of dirty paper coding (DPC). In this paper, we propose a new matrix decomposition, called the block diagonal geometric mean decomposition (BD-GMD), and develop transceiver designs that combine DPC with BD- GMD for MIMO broadcast channels. We also extend the BD- GMD to the block diagonal uniform channel decomposition (BD- UCD) with which the MIMO broadcast channel capacity can be achieved. Our proposed schemes decompose each user's MIMO channel into parallel subchannels with identical SNRs/SINRs, thus equal-rate coding can be applied across the subchannels of each user. Numerical simulations show that the proposed schemes demonstrate superior performance over conventional schemes.

Variable-rate coding for a multiple input multiple output (MIMO) channel increases transceiver complexity. In single user case, the uniform channel decomposition (UCD) scheme overcomes this problem by generating decoupled subchannels with identical SNRs so that equal-rate (ER) coding can be applied. In this paper, the solution is extended to the multi-user broadcast case. The block-diagonal geometric mean decomposition (BD-GMD) is used to design a capacity-achieving scheme, called the block-diagonal UCD (BD-UCD). It allows each user to apply ER coding on its own subchannels. An efficient near-optimal algorithm for multiuser uplink beamforming with SINR constraints is also proposed. Using this and duality, a scheme that allows ER coding to be applied to every subchannel of every user is designed. Simulations have shown that both schemes have superior BER performance and higher achievable sum-rates than conventional schemes.

Matrix decompositions play an important role in analyzing the capacity and designing the transceiver for multiple input multiple output (MIMO) channels. In the single user case, by relying on the decision feedback equalizer (DFE) at the receiver or Tomlinson-Harashima precoding (THP) at the transmitter, the geometric mean decomposition (GMD) can be used to create identical signal-to-noise ratios (SNR) for each decoupled subchannel. In this paper, we propose a new matrix decomposition, called the block-diagonal GMD (BD-GMD), for the multiuser MIMO broadcast channel. Applying THP at transmitter and linear equalization in each of the receivers, each user can achieve identical SNRs for its subchannels, thus equal-rate coding can be applied for each user. Furthermore, by using transmit power control and the BD-GMD, we design a scheme that achieves equal-rate coding for the subchannels of all users. Computer simulations have shown that the proposed schemes have better BER performances than zero-forcing THP (ZF-THP) and equal-rate ZF-THP schemes.