arXiv subjects

Mehul Motani

Publications and source records attributed to Mehul Motani.

At least 37 records · Page 2Linked to original sources

arXiv2022

Towards Better Long-range Time Series Forecasting using Generative Adversarial Networks

Long-range time series forecasting is usually based on one of two existing forecasting strategies: Direct Forecasting and Iterative Forecasting, where the former provides low bias, high variance forecasts and the later leads to low variance, high bias forecasts. In this paper, we propose a new forecasting strategy called Generative Forecasting (GenF), which generates synthetic data for the next few time steps and then makes long-range forecasts based on generated and observed data. We theoretically prove that GenF is able to better balance the forecasting variance and bias, leading to a much smaller forecasting error. We implement GenF via three components: (i) a novel conditional Wasserstein Generative Adversarial Network (GAN) based generator for synthetic time series data generation, called CWGAN-TS. (ii) a transformer based predictor, which makes long-range predictions using both generated and observed data. (iii) an information theoretic clustering algorithm to improve the training of both the CWGAN-TS and the transformer based predictor. The experimental results on five public datasets demonstrate that GenF significantly outperforms a diverse range of state-of-the-art benchmarks and classical approaches. Specifically, we find a 5% - 11% improvement in predictive performance (mean absolute error) while having a 15% - 50% reduction in parameters compared to the benchmarks. Lastly, we conduct an ablation study to demonstrate the effectiveness of the components comprising GenF.

cs.LG↗

arXiv2022

Brick Tic-Tac-Toe: Exploring the Generalizability of AlphaZero to Novel Test Environments

Traditional reinforcement learning (RL) environments typically are the same for both the training and testing phases. Hence, current RL methods are largely not generalizable to a test environment which is conceptually similar but different from what the method has been trained on, which we term the novel test environment. As an effort to push RL research towards algorithms which can generalize to novel test environments, we introduce the Brick Tic-Tac-Toe (BTTT) test bed, where the brick position in the test environment is different from that in the training environment. Using a round-robin tournament on the BTTT environment, we show that traditional RL state-search approaches such as Monte Carlo Tree Search (MCTS) and Minimax are more generalizable to novel test environments than AlphaZero is. This is surprising because AlphaZero has been shown to achieve superhuman performance in environments such as Go, Chess and Shogi, which may lead one to think that it performs well in novel test environments. Our results show that BTTT, though simple, is rich enough to explore the generalizability of AlphaZero. We find that merely increasing MCTS lookahead iterations was insufficient for AlphaZero to generalize to some novel test environments. Rather, increasing the variety of training environments helps to progressively improve generalizability across all possible starting brick configurations.

cs.LG↗

arXiv2022

DropNet: Reducing Neural Network Complexity via Iterative Pruning

Modern deep neural networks require a significant amount of computing time and power to train and deploy, which limits their usage on edge devices. Inspired by the iterative weight pruning in the Lottery Ticket Hypothesis, we propose DropNet, an iterative pruning method which prunes nodes/filters to reduce network complexity. DropNet iteratively removes nodes/filters with the lowest average post-activation value across all training samples. Empirically, we show that DropNet is robust across diverse scenarios, including MLPs and CNNs using the MNIST, CIFAR-10 and Tiny ImageNet datasets. We show that up to 90% of the nodes/filters can be removed without any significant loss of accuracy. The final pruned network performs well even with reinitialization of the weights and biases. DropNet also has similar accuracy to an oracle which greedily removes nodes/filters one at a time to minimise training loss, highlighting its effectiveness.

cs.LG↗

arXiv2021

S-Cyc: A Learning Rate Schedule for Iterative Pruning of ReLU-based Networks

We explore a new perspective on adapting the learning rate (LR) schedule to improve the performance of the ReLU-based network as it is iteratively pruned. Our work and contribution consist of four parts: (i) We find that, as the ReLU-based network is iteratively pruned, the distribution of weight gradients tends to become narrower. This leads to the finding that as the network becomes more sparse, a larger value of LR should be used to train the pruned network. (ii) Motivated by this finding, we propose a novel LR schedule, called S-Cyclical (S-Cyc) which adapts the conventional cyclical LR schedule by gradually increasing the LR upper bound (max_lr) in an S-shape as the network is iteratively pruned.We highlight that S-Cyc is a method agnostic LR schedule that applies to many iterative pruning methods. (iii) We evaluate the performance of the proposed S-Cyc and compare it to four LR schedule benchmarks. Our experimental results on three state-of-the-art networks (e.g., VGG-19, ResNet-20, ResNet-50) and two popular datasets (e.g., CIFAR-10, ImageNet-200) demonstrate that S-Cyc consistently outperforms the best performing benchmark with an improvement of 2.1% - 3.4%, without substantial increase in complexity. (iv) We evaluate S-Cyc against an oracle and show that S-Cyc achieves comparable performance to the oracle, which carefully tunes max_lr via grid search.

cs.LG↗

arXiv2019

Investigating Convolutional Neural Networks using Spatial Orderness

Convolutional Neural Networks (CNN) have been pivotal to the success of many state-of-the-art classification problems, in a wide variety of domains (for e.g. vision, speech, graphs and medical imaging). A commonality within those domains is the presence of hierarchical, spatially agglomerative local-to-global interactions within the data. For two-dimensional images, such interactions may induce an a priori relationship between the pixel data and the underlying spatial ordering of the pixels. For instance in natural images, neighboring pixels are more likely contain similar values than non-neighboring pixels which are further apart. To that end, we propose a statistical metric called spatial orderness, which quantifies the extent to which the input data (2D) obeys the underlying spatial ordering at various scales. In our experiments, we mainly find that adding convolutional layers to a CNN could be counterproductive for data bereft of spatial order at higher scales. We also observe, quite counter-intuitively, that the spatial orderness of CNN feature maps show a synchronized increase during the intial stages of training, and validation performance only improves after spatial orderness of feature maps start decreasing. Lastly, we present a theoretical analysis (and empirical validation) of the spatial orderness of network weights, where we find that using smaller kernel sizes leads to kernels of greater spatial orderness and vice-versa.

cs.LG↗

arXiv2019

Throughput Maximization with an Average Age of Information Constraint in Fading Channels

In the emerging fifth generation (5G) technology, communication nodes are expected to support two crucial classes of information traffic, namely, the enhanced mobile broadband (eMBB) traffic with high data rate requirements, and ultra-reliable low-latency communications (URLLC) traffic with strict requirements on latency and reliability. The URLLC traffic, which is usually analyzed by a metric called the age of information (AoI), is assigned the first priority over the resources at a node. Motivated by this, we consider long-term average throughput maximization problems subject to average AoI and power constraints in a single user fading channel, when (i) perfect and (ii) no channel state information at the transmitter (CSIT) is available. We propose simple age-independent stationary randomized policies (AI-SRP), which allocate powers at the transmitter based only on the channel state and/or distribution information, without any knowledge of the AoI. We show that the optimal throughputs achieved by the AI-SRPs for scenarios (i) and (ii) are at least equal to the half of the respective optimal long-term average throughputs, independent of all the parameters of the problem, and that they are within additive gaps, expressed in terms of the optimal dual variable corresponding to their average AoI constraints, from the respective optimal long-term average throughputs.

cs.IT↗

arXiv2019

Long-range Prediction of Vital Signs Using Generative Boosting via LSTM Networks

Vital signs including heart rate, respiratory rate, body temperature and blood pressure, are critical in the clinical decision making process. Effective early prediction of vital signs help to alert medical practitioner ahead of time and may prevent adverse health outcomes. In this paper, we suggest a new approach called generative boosting, in order to effectively perform early prediction of vital signs. Generative boosting consists of a generative model, to generate synthetic data for next few time steps, and several predictive models, to directly make long-range predictions based on observed and generated data. We explore generative boosting via long short-term memory (LSTM) for both the predictive and generative models, leading to a scheme called generative LSTM (GLSTM). Our experiments indicate that GLSTM outperforms a diverse range of strong benchmark models, with and without generative boosting. Finally, we use a mutual information based clustering algorithm to select a more representative dataset to train the generative model of GLSTM. This significantly improves the long-range predictive performance of high variation vital signs such as heart rate and systolic blood pressure.

cs.LG↗

arXiv2019

Multicasting Energy and Information Simultaneously

Communication systems for multicasting information and energy simultaneously to more than one user are investigated. In the system under study, a transmitter sends the same message and signal to multiple receivers over distinct and independent channels. The fundamental communication limit under a received energy constraint, called the multicast capacity-energy function, is studied and a single-letter expression is derived. This is based on coding theorems for compound channels. The problem of receiver segmentation, where receivers are divided into related groups, is also considered.

cs.IT↗

arXiv2019

Generalized Sphere-Packing Bound for Subblock-Constrained Codes

We apply the generalized sphere-packing bound to two classes of subblock-constrained codes. A la Fazeli et al. (2015), we made use of automorphism to significantly reduce the number of variables in the associated linear programming problem. In particular, we study binary constant subblock-composition codes (CSCCs), characterized by the property that the number of ones in each subblock is constant, and binary subblock energy-constrained codes (SECCs), characterized by the property that the number of ones in each subblock exceeds a certain threshold. For CSCCs, we show that the optimization problem is equivalent to finding the minimum of $N$ variables, where $N$ is independent of the number of subblocks. We then provide closed-form solutions for the generalized sphere-packing bounds for single- and double-error correcting CSCCs. For SECCs, we provide closed-form solutions for the generalized sphere-packing bounds for single errors in certain special cases. We also obtain improved bounds on the optimal asymptotic rate for CSCCs and SECCs, and provide numerical examples to highlight the improvement.

cs.IT↗

arXiv2018

Energy Harvesting Communications Using Dual Alternating Batteries

Practical energy harvesting (EH) based communication systems typically use a battery to temporarily store the harvested energy prior to its use for communication. The batteries can be damaged when they are repeatedly charged (discharged) after being partially discharged (charged), overcharged or deeply discharged. This motivates the cycle constraint which says that a battery must be charged (discharged) only after it is sufficiently discharged (charged). We also assume Bernoulli energy arrivals, and a half-duplex constraint due to which the batteries are not charged and discharged simultaneously. In this context, we study EH communication systems with: (a) a single-battery with capacity 2B units and (b) dual-batteries, each having capacity of B units. The aim is to obtain the best possible long-term average throughputs and throughput regions in point-to-point (P2P) channels and multiple access channels (MAC), respectively. For the P2P channel, we obtain an analytical optimal solution in the single-battery case, and propose optimal and sub-optimal power allocation policies for the dual-battery case. We extend these policies to obtain achievable throughput regions in MACs by jointly allocating rates and powers. From numerical simulations, we find that the optimal throughput in the dual-battery case is significantly higher than that in the single-battery case, although the total storage capacity in both cases is 2B units. Further, in the proposed policies, the largest throughput region in the single-battery case is contained within that of the dual-battery case.

cs.IT↗

arXiv2018

Second-Order Asymptotically Optimal Statistical Classification

Motivated by real-world machine learning applications, we analyze approximations to the non-asymptotic fundamental limits of statistical classification. In the binary version of this problem, given two training sequences generated according to two {\em unknown} distributions $P_1$ and $P_2$, one is tasked to classify a test sequence which is known to be generated according to either $P_1$ or $P_2$. This problem can be thought of as an analogue of the binary hypothesis testing problem but in the present setting, the generating distributions are unknown. Due to finite sample considerations, we consider the second-order asymptotics (or dispersion-type) tradeoff between type-I and type-II error probabilities for tests which ensure that (i) the type-I error probability for {\em all} pairs of distributions decays exponentially fast and (ii) the type-II error probability for a {\em particular} pair of distributions is non-vanishing. We generalize our results to classification of multiple hypotheses with the rejection option.

cs.IT↗

arXiv2018

Feature Selection Based on Unique Relevant Information for Health Data

Feature selection, which searches for the most representative features in observed data, is critical for health data analysis. Unlike feature extraction, such as PCA and autoencoder based methods, feature selection preserves interpretability, meaning that the selected features provide direct information about certain health conditions (i.e., the label). Thus, feature selection allows domain experts, such as clinicians, to understand the predictions made by machine learning based systems, as well as improve their own diagnostic skills. Mutual information is often used as a basis for feature selection since it measures dependencies between features and labels. In this paper, we introduce a novel mutual information based feature selection (MIBFS) method called SURI, which boosts features with high unique relevant information. We compare SURI to existing MIBFS methods using 3 different classifiers on 6 publicly available healthcare data sets. The results indicate that, in addition to preserving interpretability, SURI selects more relevant feature subsets which lead to higher classification performance. More importantly, we explore the dynamics of mutual information on a public low-dimensional health data set via exhaustive search. The results suggest the important role of unique relevant information in feature selection and verify the principles behind SURI.

cs.LG↗

arXiv2018

Refined Asymptotics for Rate-Distortion using Gaussian Codebooks for Arbitrary Sources

The rate-distortion saddle-point problem considered by Lapidoth (1997) consists in finding the minimum rate to compress an arbitrary ergodic source when one is constrained to use a random Gaussian codebook and minimum (Euclidean) distance encoding is employed. We extend Lapidoth's analysis in several directions in this paper. Firstly, we consider refined asymptotics. In particular, when the source is stationary and memoryless, we establish the second-order, moderate, and large deviation asymptotics of the problem. Secondly, by "random Gaussian codebook", Lapidoth referred to a collection of random codewords, each of which is drawn independently and uniformly from the surface of an $n$-dimensional sphere. To be more precise, we term this as a spherical codebook. We also consider i.i.d.\ Gaussian codebooks in which each random codeword is drawn independently from a product Gaussian distribution. We derive the second-order, moderate, and large deviation asymptotics when i.i.d.\ Gaussian codebooks are employed. In contrast to the recent work on the channel coding counterpart by Scarlett, Tan and Durisi (2017), the dispersions for spherical and i.i.d.\ Gaussian codebooks are identical. The ensemble excess-distortion exponents for both spherical and i.i.d.\ Gaussian codebooks are established for all rates. Furthermore, we show that the i.i.d.\ Gaussian codebook has a strictly larger excess-distortion exponent than its spherical counterpart for any rate greater than the ensemble rate-distortion function derived by Lapidoth.

cs.IT↗

arXiv2018

The Dispersion of Mismatched Joint Source-Channel Coding for Arbitrary Sources and Additive Channels

We consider a joint source channel coding (JSCC) problem in which we desire to transmit an arbitrary memoryless source over an arbitrary additive channel. We propose a mismatched coding architecture that consists of Gaussian codebooks for both the source reproduction sequences and channel codewords. The natural nearest neighbor encoder and decoder, however, need to be judiciously modified to obtain the highest communication rates at finite blocklength. In particular, we consider a unequal error protection (UEP) scheme in which all sources are partitioned into disjoint power type classes. We also regularize the nearest neighbor decoder so that an appropriate measure of the size of each power type class is taken into account in the decoding strategy. For such an architecture, we derive ensemble-tight second-order and moderate deviations results. Our first-order (optimal bandwidth expansion ratio) result generalizes the seminal results by Lapidoth (1996, 1997). The dispersion of our JSCC scheme is a linear combination of the mismatched dispersions for the channel coding saddle-point problem by Scarlett, Tan and Durisi (2017) and the rate-distortion saddle-point problem by the present authors, thus also generalizing these results.

cs.IT↗

arXiv2018

Are RLL Codes Suitable for Simultaneous Energy and Information Transfer?

Run-length limited (RLL) codes are a well-studied class of constrained codes having application in diverse areas such as optical and magnetic data recording systems, DNA-based storage, and visible light communication. RLL codes have also been proposed for the emerging area of simultaneous energy and information transfer, where the receiver uses the received signal for decoding information as well as for harvesting energy to run its circuitry. In this paper, we show that RLL codes are not the best codes for simultaneous energy and information transfer, in terms of the maximum number of codewords which avoid energy outage, i.e., outage-constrained capacity. Specifically, we show that sliding window constrained (SWC) codes and subblock energy constrained (SEC) codes have significantly higher outage-constrained capacities than RLL codes.

cs.IT↗

arXiv2018

Hybrid NOMA-TDMA for Multiple Access Channels with Non-Ideal Batteries and Circuit Cost

We consider a multiple-access channel where the users are powered from batteries having non-negligible internal resistance. When power is drawn from the battery, a variable fraction of the power, which is a function of the power drawn from the battery, is lost across the internal resistance. Hence, the power delivered to the load is less than the power drawn from the battery. The users consume a constant power for the circuit operation during transmission but do not consume any power when not transmitting. In this setting, we obtain the maximum sum-rates and achievable rate regions under various cases. We show that, unlike in the ideal battery case, the TDMA (time-division multiple access) strategy, wherein the users transmit orthogonally in time, may not always achieve the maximum sum-rate when the internal resistance is non-zero. The users may need to adopt a hybrid NOMA-TDMA strategy which combines the features of NOMA (non-orthogonal multiple access) and TDMA, wherein a set of users are allocated fixed time windows for orthogonal single-user and non-orthogonal joint transmissions, respectively. We also numerically show that the maximum achievable rate regions in NOMA and TDMA strategies are contained within the maximum achievable rate region of the hybrid NOMA-TDMA strategy.

cs.IT↗

arXiv2018

Skip-Sliding Window Codes

Constrained coding is used widely in digital communication and storage systems. In this paper, we study a generalized sliding window constraint called the skip-sliding window. A skip-sliding window (SSW) code is defined in terms of the length $L$ of a sliding window, skip length $J$, and cost constraint $E$ in each sliding window. Each valid codeword of length $L + kJ$ is determined by $k+1$ windows of length $L$ where window $i$ starts at $(iJ + 1)$th symbol for all non-negative integers $i$ such that $i \leq k$; and the cost constraint $E$ in each window must be satisfied. In this work, two methods are given to enumerate the size of SSW codes and further refinements are made to reduce the enumeration complexity. Using the proposed enumeration methods, the noiseless capacity of binary SSW codes is determined and observations such as greater capacity than other classes of codes are made. Moreover, some noisy capacity bounds are given. SSW coding constraints arise in various applications including simultaneous energy and information transfer.

cs.IT↗

arXiv2017

Exponential Strong Converse for Content Identification with Lossy Recovery

We revisit the high-dimensional content identification with lossy recovery problem (Tuncel and Gündüz, 2014) and establish an exponential strong converse theorem. As a corollary of the exponential strong converse theorem, we derive an upper bound on the joint identification-error and excess-distortion exponent for the problem. Our main results can be specialized to the biometrical identification problem~(Willems, 2003) and the content identification problem~(Tuncel, 2009) since these two problems are both special cases of the content identification with lossy recovery problem. We leverage the information spectrum method introduced by Oohama and adapt the strong converse techniques therein to be applicable to the problem at hand.

cs.IT↗