A DRL Approach for RIS-Assisted Full-Duplex UL and DL Transmission: Beamforming, Phase Shift and Power Optimization

In this work, a two-stage deep reinforcement learning (DRL) approach is presented for a full-duplex (FD) transmission scenario that does not depend on the channel state information (CSI) knowledge to predict the phase-shifts of reconfigurable intelligent surface (RIS), beamformers at the base station (BS), and the transmit powers of BS and uplink users in order to maximize the weighted sum rate of uplink and downlink users. As the self-interference (SI) cancellation and beamformer design are coupled problems, the first stage uses a least squares method to partially cancel self-interference (SI) and initiate learning, while the second stage uses DRL to make predictions and achieve performance close to methods with perfect CSI knowledge. Further, to reduce the signaling from BS to the RISs, a DRL framework is proposed that predicts quantized RIS phase-shifts and beamformers using $32$ times fewer bits than the continuous version. The quantized methods have reduced action space and therefore faster convergence; with sufficient training, the UL and DL rates for the quantized phase method are $8.14\%$ and $2.45\%$ better than the continuous phase method respectively. The RIS elements can be grouped to have similar phase-shifts to further reduce signaling, at the cost of reduced performance.

Rotate the ReLU to implicitly sparsify deep networks

In the era of Deep Neural Network based solutions for a variety of real-life tasks, having a compact and energy-efficient deployable model has become fairly important. Most of the existing deep architectures use Rectifier Linear Unit (ReLU) activation. In this paper, we propose a novel idea of rotating the ReLU activation to give one more degree of freedom to the architecture. We show that this activation wherein the rotation is learned via training results in the elimination of those parameters/filters in the network which are not important for the task. In other words, rotated ReLU seems to be doing implicit sparsification. The slopes of the rotated ReLU activations act as coarse feature extractors and unnecessary features can be eliminated before retraining. Our studies indicate that features always choose to pass through a lesser number of filters in architectures such as ResNet and its variants. Hence, by rotating the ReLU, the weights or the filters that are not necessary are automatically identified and can be dropped thus giving rise to significant savings in memory and computation. Furthermore, in some cases, we also notice that along with saving in memory and computation we also obtain improvement over the reported performance of the corresponding baseline work in the popular datasets such as MNIST, CIFAR-10, CIFAR-100, and SVHN.

Binarized ResNet: Enabling Automatic Modulation Classification at the resource-constrained Edge

In this paper, we propose a ResNet based neural architecture to solve the problem of Automatic Modulation Classification. We showed that our architecture outperforms the state-of-the-art (SOTA) architectures. We further propose to binarize the network to deploy it in the Edge network where the devices are resource-constrained i.e. have limited memory and computing power. Instead of simple binarization, rotated binarization is applied to the network which helps to close the significant performance gap between the real and the binarized network. Because of the immense representation capability or the real network, its rotated binarized version achieves $85.33\%$ accuracy compared to $95.76\%$ accuracy of the proposed real network with $2.33$ and $16$ times lesser computing power than two of the SOTA architectures, MCNet and RMLResNet respectively, and approximately $16$ times less memory than both. The performance can be improved further to $87.74\%$ by taking an ensemble of four such rotated binarized networks.

BayesAoA: A Bayesian method for Computation Efficient Angle of Arrival Estimation

The angle of Arrival (AoA) estimation is of great interest in modern communication systems. Traditional maximum likelihood-based iterative algorithms are sensitive to initialization and cannot be used online. We propose a Bayesian method to find AoA that is insensitive towards initialization. The proposed method is less complex and needs fewer computing resources than traditional deep learning-based methods. It has a faster convergence than the brute-force methods. Further, a Hedge type solution is proposed that helps to deploy the method online to handle the situations where the channel noise and antenna configuration in the receiver change over time. The proposed method achieves $92\%$ accuracy in a channel of noise variance $10^{-6}$ with $19.3\%$ of the brute-force method's computation.

Realizing Neural Decoder at the Edge with Ensembled BNN

In this work, we propose extreme compression techniques like binarization, ternarization for Neural Decoders such as TurboAE. These methods reduce memory and computation by a factor of 64 with a performance better than the quantized (with 1-bit or 2-bits) Neural Decoders. However, because of the limited representation capability of the Binary and Ternary networks, the performance is not as good as the real-valued decoder. To fill this gap, we further propose to ensemble 4 such weak performers to deploy in the edge to achieve a performance similar to the real-valued network. These ensemble decoders give 16 and 64 times saving in memory and computation respectively and help to achieve performance similar to real-valued TurboAE.

Understanding Learning Dynamics of Binary Neural Networks via Information Bottleneck

Compact neural networks are essential for affordable and power efficient deep learning solutions. Binary Neural Networks (BNNs) take compactification to the extreme by constraining both weights and activations to two levels, $\{+1, -1\}$. However, training BNNs are not easy due to the discontinuity in activation functions, and the training dynamics of BNNs is not well understood. In this paper, we present an information-theoretic perspective of BNN training. We analyze BNNs through the Information Bottleneck principle and observe that the training dynamics of BNNs is considerably different from that of Deep Neural Networks (DNNs). While DNNs have a separate empirical risk minimization and representation compression phases, our numerical experiments show that in BNNs, both these phases are simultaneous. Since BNNs have a less expressive capacity, they tend to find efficient hidden representations concurrently with label fitting. Experiments in multiple datasets support these observations, and we see a consistent behavior across different activation functions in BNNs.

What is the optimal depth for deep-unfolding architectures at deployment?

Recently, many iterative algorithms proposed for various applications such as compressed sensing, MIMO Detection, etc. have been unfolded and presented as deep networks; these networks are shown to produce better results than the algorithms in their iterative forms. However, deep networks are highly sensitive to the hyperparameters chosen. Especially for a deep unfolded network, using more layers may lead to redundancy and hence, excessive computation during deployment. In this work, we consider the problem of determining the optimal number of layers required for such unfolded architectures. We propose a method that treats the networks as experts and measures the relative importance of the expertise provided by layers using a variant of the popular Hedge algorithm. Based on the importance of the different layers, we determine the optimal layers required for deployment. We study the effectiveness of this method by applying it to two recent and popular deep-unfolding architectures, namely DetNet and TISTANet.