arXiv subjects

Sriram Vishwanath

Publications and source records attributed to Sriram Vishwanath.

At least 73 records · Page 4Linked to original sources

arXiv2020

Long Short-Term Memory Spiking Networks and Their Applications

Recent advances in event-based neuromorphic systems have resulted in significant interest in the use and development of spiking neural networks (SNNs). However, the non-differentiable nature of spiking neurons makes SNNs incompatible with conventional backpropagation techniques. In spite of the significant progress made in training conventional deep neural networks (DNNs), training methods for SNNs still remain relatively poorly understood. In this paper, we present a novel framework for training recurrent SNNs. Analogous to the benefits presented by recurrent neural networks (RNNs) in learning time series models within DNNs, we develop SNNs based on long short-term memory (LSTM) networks. We show that LSTM spiking networks learn the timing of the spikes and temporal dependencies. We also develop a methodology for error backpropagation within LSTM-based SNNs. The developed architecture and method for backpropagation within LSTM-based SNNs enable them to learn long-term dependencies with comparable results to conventional LSTMs.

cs.NE↗

arXiv2020

Beamforming Cancellation Design for Millimeter-Wave Full-Duplex

In recent years, there has been extensive research on millimeter-wave (mmWave) communication and on in-band full-duplex (FD) communication, but work on the combination of the two is relatively lacking. FD mmWave systems could offer increased spectral efficiency and decreased latency while also suggesting the redesign of existing mmWave applications. While FD technology has been well-explored for sub-6 GHz systems, the developed methods do not translate well to mmWave. This turns us to a method called beamforming cancellation (BFC), where the highly directional mmWave beams are steered to mitigate self-interference (SI) and enable simultaneous transmission and reception in-band. In this paper, we present BFC designs for two fully-connected hybrid beamforming scenarios, both of which sufficiently suppress the SI such that the sum spectral efficiency approaches that of a SI-free FD system. A simulation and its results are then used to verify our designs.

eess.SP↗

arXiv2020

Deep Networks as Logical Circuits: Generalization and Interpretation

Not only are Deep Neural Networks (DNNs) black box models, but also we frequently conceptualize them as such. We lack good interpretations of the mechanisms linking inputs to outputs. Therefore, we find it difficult to analyze in human-meaningful terms (1) what the network learned and (2) whether the network learned. We present a hierarchical decomposition of the DNN discrete classification map into logical (AND/OR) combinations of intermediate (True/False) classifiers of the input. Those classifiers that can not be further decomposed, called atoms, are (interpretable) linear classifiers. Taken together, we obtain a logical circuit with linear classifier inputs that computes the same label as the DNN. This circuit does not structurally resemble the network architecture, and it may require many fewer parameters, depending on the configuration of weights. In these cases, we obtain simultaneously an interpretation and generalization bound (for the original DNN), connecting two fronts which have historically been investigated separately. Unlike compression techniques, our representation is. We motivate the utility of this perspective by studying DNNs in simple, controlled settings, where we obtain superior generalization bounds despite using only combinatorial information (e.g. no margin information). We demonstrate how to "open the black box" on the MNIST dataset. We show that the learned, internal, logical computations correspond to semantically meaningful (unlabeled) categories that allow DNN descriptions in plain English. We improve the generalization of an already trained network by interpreting, diagnosing, and replacing components the logical circuit that is the DNN.

cs.LG↗

arXiv2020

Interpretable Factorization for Neural Network ECG Models

The ability of deep learning (DL) to improve the practice of medicine and its clinical outcomes faces a looming obstacle: model interpretation. Without description of how outputs are generated, a collaborating physician can neither resolve when the model's conclusions are in conflict with his or her own, nor learn to anticipate model behavior. Current research aims to interpret networks that diagnose ECG recordings, which has great potential impact as recordings become more personalized and widely deployed. A generalizable impact beyond ECGs lies in the ability to provide a rich test-bed for the development of interpretive techniques in medicine. Interpretive techniques for Deep Neural Networks (DNNs), however, tend to be heuristic and observational in nature, lacking the mathematical rigor one might expect in the analysis of math equations. The motivation of this paper is to offer a third option, a scientific approach. We treat the model output itself as a phenomenon to be explained through component parts and equations governing their behavior. We argue that these component parts should also be "black boxes" --additional targets to interpret heuristically with clear functional connection to the original. We show how to rigorously factor a DNN into a hierarchical equation consisting of black box variables. This is not a subdivision into physical parts, like an organism into its cells; it is but one choice of an equation into a collection of abstract functions. Yet, for DNNs trained to identify normal ECG waveforms on PhysioNet 2017 Challenge data, we demonstrate this choice yields interpretable component models identified with visual composite sketches of ECG samples in corresponding input regions. Moreover, the recursion distills this interpretation: additional factorization of component black boxes corresponds to ECG partitions that are more morphologically pure.

cs.LG↗

arXiv2020

Robust Face Verification via Disentangled Representations

We introduce a robust algorithm for face verification, i.e., deciding whether twoimages are of the same person or not. Our approach is a novel take on the idea ofusing deep generative networks for adversarial robustness. We use the generativemodel during training as an online augmentation method instead of a test-timepurifier that removes adversarial noise. Our architecture uses a contrastive loss termand a disentangled generative model to sample negative pairs. Instead of randomlypairing two real images, we pair an image with its class-modified counterpart whilekeeping its content (pose, head tilt, hair, etc.) intact. This enables us to efficientlysample hard negative pairs for the contrastive loss. We experimentally show that, when coupled with adversarial training, the proposed scheme converges with aweak inner solver and has a higher clean and robust accuracy than state-of-the-art-methods when evaluated against white-box physical attacks.

cs.CV↗

arXiv2020

Model-free Reinforcement Learning for Stochastic Stackelberg Security Games

In this paper, we consider a sequential stochastic Stackelberg game with two players, a leader and a follower. The follower has access to the state of the system while the leader does not. Assuming that the players act in their respective best interests, the follower's strategy is to play the best response to the leader's strategy. In such a scenario, the leader has the advantage of committing to a policy which maximizes its own returns given the knowledge that the follower is going to play the best response to its policy. Thus, both players converge to a pair of policies that form the Stackelberg equilibrium of the game. Recently,~[1] provided a sequential decomposition algorithm to compute the Stackelberg equilibrium for such games which allow for the computation of Markovian equilibrium policies in linear time as opposed to double exponential, as before. In this paper, we extend the idea to an MDP whose dynamics are not known to the players, to propose an RL algorithm based on Expected Sarsa that learns the Stackelberg equilibrium policy by simulating a model of the MDP. We use particle filters to estimate the belief update for a common agent which computes the optimal policy based on the information which is common to both the players. We present a security game example to illustrate the policy learned by our algorithm. by simulating a model of the MDP. We use particle filters to estimate the belief update for a common agent which computes the optimal policy based on the information which is common to both the players. We present a security game example to illustrate the policy learned by our algorithm.

math.OC↗

arXiv2020

Model-free Reinforcement Learning for Non-stationary Mean Field Games

In this paper, we consider a finite horizon, non-stationary, mean field games (MFG) with a large population of homogeneous players, sequentially making strategic decisions, where each player is affected by other players through an aggregate population state termed as mean field state. Each player has a private type that only it can observe, and a mean field population state representing the empirical distribution of other players' types, which is shared among all of them. Recently, authors in [1] provided a sequential decomposition algorithm to compute mean field equilibrium (MFE) for such games which allows for the computation of equilibrium policies for them in linear time than exponential, as before. In this paper, we extend it for the case when state transitions are not known, to propose a reinforcement learning algorithm based on Expected Sarsa with a policy gradient approach that learns the MFE policy by learning the dynamics of the game simultaneously. We illustrate our results using cyber-physical security example.

eess.SY↗

arXiv2020

Decentralized multi-agent reinforcement learning with shared actions

In this paper, we propose a novel model-free reinforcement learning algorithm to compute the optimal policies for a multi-agent system with $N$ cooperative agents where each agent privately observes it's own private type and publicly observes each others' actions. The goal is to maximize their collective reward. The problem belongs to the broad class of decentralized control problems with partial information. We use the common agent approach wherein some fictitious common agent picks the best policy based on a belief on the current states of the agents. These beliefs are updated individually for each agent from their current belief and action histories. Belief state updates without the knowledge of system dynamics is a challenge. In this paper, we employ particle filters called the bootstrap filter distributively across agents to update the belief. We provide a model-free reinforcement learning (RL) method for this multi-agent partially observable Markov decision processes using the particle filter and sampled trajectories to estimate the optimal policies for the agents. We showcase our results with the help of a smartgrid application where the users strive to reduce collective cost of power for all the agents in the grid. Finally, we compare the performances for model and model-free implementation of the RL algorithm establishing the effectiveness of particle filter (pf) method.

eess.SY↗

arXiv2020

Sample Compression, Support Vectors, and Generalization in Deep Learning

Even though Deep Neural Networks (DNNs) are widely celebrated for their practical performance, they possess many intriguing properties related to depth that are difficult to explain both theoretically and intuitively. Understanding how weights in deep networks coordinate together across layers to form useful learners has proven challenging, in part because the repeated composition of nonlinearities has proved intractable. This paper presents a reparameterization of DNNs as a linear function of a feature map that is locally independent of the weights. This feature map transforms depth-dependencies into simple tensor products and maps each input to a discrete subset of the feature space. Then, using a max-margin assumption, the paper develops a sample compression representation of the neural network in terms of the discrete activation state of neurons induced by s ``support vectors". The paper shows that the number of support vectors s relates with learning guarantees for neural networks through sample compression bounds, yielding a sample complexity of O(ns/epsilon) for networks with n neurons. Finally, the number of support vectors s is found to monotonically increase with width and label noise but decrease with depth.

cs.LG↗

arXiv2020

Detecting Patch Adversarial Attacks with Image Residuals

We introduce an adversarial sample detection algorithm based on image residuals, specifically designed to guard against patch-based attacks. The image residual is obtained as the difference between an input image and a denoised version of it, and a discriminator is trained to distinguish between clean and adversarial samples. More precisely, we use a wavelet domain algorithm for denoising images and demonstrate that the obtained residuals act as a digital fingerprint for adversarial attacks. To emulate the limitations of a physical adversary, we evaluate the performance of our approach against localized (patch-based) adversarial attacks, including in settings where the adversary has complete knowledge about the detection scheme. Our results show that the proposed detection method generalizes to previously unseen, stronger attacks and that it is able to reduce the success rate (conversely, increase the computational effort) of an adaptive attacker.

cs.CV↗

arXiv2020

Equipping Millimeter-Wave Full-Duplex with Analog Self-Interference Cancellation

There have been recent works on enabling in-band full-duplex operation using millimeter-wave (mmWave) transceivers. These works are based solely on creating sufficient isolation between a transceiver's transmitter and receiver via multiple-input multiple-output (MIMO) precoding and combining. In this work, we propose supplementing these beamforming strategies with analog self-interference cancellation (A-SIC). By leveraging A-SIC, a portion of the self-interference is cancelled without the need for beamforming, allowing for more optimal beamforming strategies to be used in serving users. We use simulation to demonstrate that even with finite resolution A-SIC solutions, there are significant gains to be had in sum spectral efficiency. With a single bit of A-SIC resolution, improvements over a beamforming-only design are present. With 8 bits of A-SIC resolution, our design nearly approaches that of ideal full-duplex operation. To the best of our knowledge, this is the first mmWave full-duplex design that combines both beamforming and A-SIC to achieve simultaneous transmission and reception in-band.

eess.SP↗

arXiv2020

Achieving Multi-Port Memory Performance on Single-Port Memory with Coding Techniques

Many performance critical systems today must rely on performance enhancements, such as multi-port memories, to keep up with the increasing demand of memory-access capacity. However, the large area footprints and complexity of existing multi-port memory designs limit their applicability. This paper explores a coding theoretic framework to address this problem. In particular, this paper introduces a framework to encode data across multiple single-port memory banks in order to {\em algorithmically} realize the functionality of multi-port memory. This paper proposes three code designs with significantly less storage overhead compared to the existing replication based emulations of multi-port memories. To further improve performance, we also demonstrate a memory controller design that utilizes redundancy across coded memory banks to more efficiently schedule read and write requests sent across multiple cores. Furthermore, guided by DRAM traces, the paper explores {\em dynamic coding} techniques to improve the efficiency of the coding based memory design. We then show significant performance improvements in critical word read and write latency in the proposed coded-memory design when compared to a traditional uncoded-memory design.

cs.AR↗

arXiv2020

Learning Representations by Maximizing Mutual Information in Variational Autoencoders

Variational autoencoders (VAEs) have ushered in a new era of unsupervised learning methods for complex distributions. Although these techniques are elegant in their approach, they are typically not useful for representation learning. In this work, we propose a simple yet powerful class of VAEs that simultaneously result in meaningful learned representations. Our solution is to combine traditional VAEs with mutual information maximization, with the goal to enhance amortized inference in VAEs using Information Theoretic techniques. We call this approach InfoMax-VAE, and such an approach can significantly boost the quality of learned high-level representations. We realize this through the explicit maximization of information measures associated with the representation. Using extensive experiments on varied datasets and setups, we show that InfoMax-VAE outperforms contemporary popular approaches, including Info-VAE and $β$-VAE.

cs.LG↗

arXiv2019

Enabling In-Band Coexistence of Millimeter-Wave Communication and Radar

The wide bandwidths available at millimeter-wave (mmWave) frequencies have offered exciting potential to wireless communication systems and radar alike. Communication systems can offer higher rates and support more users with mmWave bands while radar systems can benefit from higher resolution captures. This leads to the possibility that portions of mmWave spectrum will be occupied by both communication and radar (e.g., 60 GHz industrial, scientific, and medical (ISM) band). This potential coexistence motivates the work of this paper, in which we present a design that can enable simultaneous, in-band operation of a communication system and radar system across the same mmWave frequencies. To enable such a feat, we mitigate the interference that would otherwise be incurred by leveraging the numerous antennas offered in mmWave communication systems. Dense antenna arrays allow us to avoid interference spatially, even with the hybrid beamforming constraints often imposed by mmWave communication systems. Simulation shows that our design sufficiently enables simultaneous, in-band coexistence of a mmWave radar and communication system.

eess.SP↗

arXiv2019

Frequency-Selective Beamforming Cancellation Design for Millimeter-Wave Full-Duplex

The wide bandwidths offered at millimeter-wave (mmWave) frequencies have made them an attractive choice for future wireless communication systems. Recent works have presented beamforming strategies for enabling in-band full-duplex (FD) capability at mmWave even under the constraints of hybrid beamforming, extending the exciting possibilities of next-generation wireless. Existing mmWave FD designs, however, do not consider frequency-selective mmWave channels. Wideband communication at mmWave suggests that frequency-selectivity will likely be of concern since communication channels will be on the order of hundreds of megahertz or more. This has motivated the work of this paper, in which we present a frequency-selective beamforming design to enable practical wideband mmWave FD applications. In our designs, we account for the challenges associated with hybrid analog/digital beamforming such as phase shifter resolution, a desirably low number of radio frequency (RF) chains, and the frequency-flat nature of analog beamformers. We use simulation to validate our work, which indicates that spectral efficiency gains can be achieved with our design by enabling simultaneous transmission and reception in-band.

eess.SP↗

arXiv2019

HashCore: Proof-of-Work Functions for General Purpose Processors

Over the past five years, the rewards associated with mining Proof-of-Work blockchains have increased substantially. As a result, miners are heavily incentivized to design and utilize Application Specific Integrated Circuits (ASICs) that can compute hashes far more efficiently than existing general purpose hardware. Currently, it is difficult for most users to purchase and operate ASICs due to pricing and availability constraints, resulting in a relatively small number of miners with respect to total user base for most popular cryptocurrencies. In this work, we aim to invert the problem of ASIC development by constructing a Proof-of-Work function for which an existing general purpose processor (GPP, such as an x86 IC) is already an optimized ASIC. In doing so, we will ensure that any would-be miner either already owns an ASIC for the Proof-of-Work system they wish to participate in or can attain one at a competitive price with relative ease. In order to achieve this, we present HashCore, a Proof-of-Work function composed of "widgets" generated pseudo-randomly at runtime that each execute a sequence of general purpose processor instructions designed to stress the computational resources of such a GPP. The widgets will be modeled after workloads that GPPs have been optimized for, for example, the SPEC CPU 2017 benchmark suite for x86 ICs, in a technique we refer to as inverted benchmarking. We provide a proof that HashCore is collision-resistant regardless of how the widgets are implemented. We observe that GPP designers/developers essentially create an ASIC for benchmarks such as SPEC CPU 2017. By modeling HashCore after such benchmarks, we create a Proof-of-Work function that can be run most efficiently on a GPP, resulting in a more accessible, competitive, and balanced mining market.

cs.CR↗

arXiv2018

Experimental Design for Cost-Aware Learning of Causal Graphs

We consider the minimum cost intervention design problem: Given the essential graph of a causal graph and a cost to intervene on a variable, identify the set of interventions with minimum total cost that can learn any causal graph with the given essential graph. We first show that this problem is NP-hard. We then prove that we can achieve a constant factor approximation to this problem with a greedy algorithm. We then constrain the sparsity of each intervention. We develop an algorithm that returns an intervention design that is nearly optimal in terms of size for sparse graphs with sparse interventions and we discuss how to use it when there are costs on the vertices.

cs.LG↗

arXiv2018

On the Key Generation Rate of Physically Unclonable Functions

In this paper, an algebraic binning based coding scheme and its associated achievable rate for key generation using physically unclonable functions (PUFs) is determined. This achievable rate is shown to be optimal under the generated-secret (GS) model for PUFs. Furthermore, a polar code based polynomial-time encoding and decoding scheme that achieves this rate is also presented.

cs.IT↗