arXiv subjects

Vwani P. Roychowdhury

Publications and source records attributed to Vwani P. Roychowdhury.

At least 19 recordsLinked to original sources

arXiv2020

Accuracy and Resiliency of Analog Compute-in-Memory Inference Engines

Recently, analog compute-in-memory (CIM) architectures based on emerging analog non-volatile memory (NVM) technologies have been explored for deep neural networks (DNN) to improve energy efficiency. Such architectures, however, leverage charge conservation, an operation with infinite resolution, and thus are susceptible to errors. The computations in DNN realized by analog NVM thus have high uncertainty due to the device stochasticity. Several reports have demonstrated the use of analog NVM for CIM in a limited scale. It is unclear whether the uncertainties in computations will prohibit large-scale DNNs. To explore this critical issue of scalability, this paper first presents a simulation framework to evaluate the feasibility of large-scale DNNs based on CIM architecture and analog NVM. Simulation results show that DNNs trained for high-precision digital computing engines are not resilient against the uncertainty of the analog NVM devices. To avoid such catastrophic failures, this paper introduces the analog floating-point representation for the DNN, and the Hessian-Aware Stochastic Gradient Descent (HA-SGD) training algorithm to enhance the inference accuracy of trained DNNs. As a result of such enhancements, DNNs such as Wide ResNets for the CIFAR-100 image recognition problem are demonstrated to have significant performance improvements in accuracy without adding cost to the inference hardware.

eess.SP↗

arXiv2013

Blind Men and the Elephant: Detecting Evolving Groups In Social News

We propose an automated and unsupervised methodology for a novel summarization of group behavior based on content preference. We show that graph theoretical community evolution (based on similarity of user preference for content) is effective in indexing these dynamics. Combined with text analysis that targets automatically-identified representative content for each community, our method produces a novel multi-layered representation of evolving group behavior. We demonstrate this methodology in the context of political discourse on a social news site with data that spans more than four years and find coexisting political leanings over extended periods and a disruptive external event that lead to a significant reorganization of existing patterns. Finally, where there exists no ground truth, we propose a new evaluation approach by using entropy measures as evidence of coherence along the evolution path of these groups. This methodology is valuable to designers and managers of online forums in need of granular analytics of user activity, as well as to researchers in social and political sciences who wish to extend their inquiries to large-scale data available on the web.

cs.SI↗

arXiv2009

Performance of Magnetic Quantum Cellular Automata and Limitations due to Thermal Noise

Operation parameters of magnetic quantum cellular automata are evaluated for the purposes of reliable logic operation. The dynamics of the nanomagnets is simulated via the Landau-Lifshitz-Gilbert equations with a stochastic magnetic field corresponding to thermal fluctuations. It is found that in the macrospin approximation the switching speed does not change under scaling of both size and distances between nanomagnets. Thermal fluctuations put a limitation on the size of nanomagnets, since the gate error rate becomes excessive for nanomagnets smaller than 200nm at room temperature.

cond-mat.mes-hall↗

arXiv2009

Experience versus Talent Shapes the Structure of the Web

We use sequential large-scale crawl data to empirically investigate and validate the dynamics that underlie the evolution of the structure of the web. We find that the overall structure of the web is defined by an intricate interplay between experience or entitlement of the pages (as measured by the number of inbound hyperlinks a page already has), inherent talent or fitness of the pages (as measured by the likelihood that someone visiting the page would give a hyperlink to it), and the continual high rates of birth and death of pages on the web. We find that the web is conservative in judging talent and the overall fitness distribution is exponential, showing low variability. The small variance in talent, however, is enough to lead to experience distributions with high variance: The preferential attachment mechanism amplifies these small biases and leads to heavy-tailed power-law (PL) inbound degree distributions over all pages, as well as over pages that are of the same age. The balancing act between experience and talent on the web allows newly introduced pages with novel and interesting content to grow quickly and surpass older pages. In this regard, it is much like what we observe in high-mobility and meritocratic societies: People with entitlement continue to have access to the best resources, but there is just enough screening for fitness that allows for talented winners to emerge and join the ranks of the leaders. Finally, we show that the fitness estimates have potential practical applications in ranking query results.

cs.CY↗

arXiv2008

To Broad-Match or Not to Broad-Match : An Auctioneer's Dilemma ?

We initiate the study of an interesting aspect of sponsored search advertising, namely the consequences of broad match-a feature where an ad of an advertiser can be mapped to a broader range of relevant queries, and not necessarily to the particular keyword(s) that ad is associated with. Starting with a very natural setting for strategies available to the advertisers, and via a careful look through the algorithmic lens, we first propose solution concepts for the game originating from the strategic behavior of advertisers as they try to optimize their budget allocation across various keywords. Next, we consider two broad match scenarios based on factors such as information asymmetry between advertisers and the auctioneer, and the extent of auctioneer's control on the budget splitting. In the first scenario, the advertisers have the full information about broad match and relevant parameters, and can reapportion their own budgets to utilize the extra information; in particular, the auctioneer has no direct control over budget splitting. We show that, the same broad match may lead to different equilibria, one leading to a revenue improvement, whereas another to a revenue loss. This leaves the auctioneer in a dilemma - whether to broad-match or not. This motivates us to consider another broad match scenario, where the advertisers have information only about the current scenario, and the allocation of the budgets unspent in the current scenario is in the control of the auctioneer. We observe that the auctioneer can always improve his revenue by judiciously using broad match. Thus, information seems to be a double-edged sword for the auctioneer.

cs.GT↗

arXiv2008

Latency in local, two-dimensional, fault-tolerant quantum computing

We analyze the latency of fault-tolerant quantum computing based on the 9-qubit Bacon-Shor code using a local, two-dimensional architecture. We embed the data qubits in a 7 by 7 array of physical qubits, where the extra qubits are used for ancilla preparation and qubit transportation by means of a SWAP chain. The latency is reduced with respect to a similar implementation using Steane's 7-qubit code (K. M. Svore, D. P. DiVincenzo, and B. M. Terhal, Quantum Information & Computation {\bf 7}, 297 (2007)). Furthermore, the error threshold is also improved to $2.02 \times 10^{-5}$, when memory errors are taken to be one tenth of the gate error rates.

quant-ph↗

arXiv2007

Preferential survival in models of complex ad hoc networks

There has been a rich interplay in recent years between (i) empirical investigations of real world dynamic networks, (ii) analytical modeling of the microscopic mechanisms that drive the emergence of such networks, and (iii) harnessing of these mechanisms to either manipulate existing networks, or engineer new networks for specific tasks. We continue in this vein, and study the deletion phenomenon in the web by following two different sets of web-sites (each comprising more than 150,000 pages) over a one-year period. Empirical data show that there is a significant deletion component in the underlying web networks, but the deletion process is not uniform. This motivates us to introduce a new mechanism of preferential survival (PS), where nodes are removed according to a degree-dependent deletion kernel. We use the mean-field rate equation approach to study a general dynamic model driven by Preferential Attachment (PA), Double PA (DPA), and a tunable PS, where c nodes (c<1) are deleted per node added to the network, and verify our predictions via large-scale simulations. One of our results shows that, unlike in the case of uniform deletion, the PS kernel when coupled with the standard PA mechanism, can lead to heavy-tailed power law networks even in the presence of extreme turnover in the network. Moreover, a weak DPA mechanism, coupled with PS, can help make the network even more heavy-tailed, especially in the limit when deletion and insertion rates are almost equal, and the overall network growth is minimal. The dynamics reported in this work can be used to design and engineer stable ad hoc networks and explain the stability of the power law exponents observed in real-world networks.

physics.soc-ph↗

arXiv2007

Capacity as a Fundamental Metric for Mechanism Design in the Information Economy

The auction theory literature has so far focused mostly on the design of mechanisms that takes the revenue or the efficiency as a yardstick. However, scenarios where the {\it capacity}, which we define as \textit{``the number of bidders the auctioneer wants to have a positive probability of getting the item''}, is a fundamental concern are ubiquitous in the information economy. For instance, in sponsored search auctions (SSA's) or in online ad-exchanges, the true value of an ad-slot for an advertiser is inherently derived from the conversion-rate, which in turn depends on whether the advertiser actually obtained the ad-slot or not; thus, unless the capacity of the underlying auction is large, key parameters, such as true valuations and advertiser-specific conversion rates, will remain unknown or uncertain leading to inherent inefficiencies in the system. In general, the same holds true for all information goods/digital goods. We initiate a study of mechanisms, which take capacity as a yardstick, in addition to revenue/efficiency. We show that in the case of a single indivisible item one simple way to incorporate capacity constraints is via designing mechanisms to sell probability distributions, and that under certain conditions, such optimal probability distributions could be identified using a Linear programming approach. We define a quantity called {\it price of capacity} to capture the tradeoff between capacity and revenue/efficiency. We also study the case of sponsored search auctions. Finally, we discuss how general such an approach via probability spikes can be made, and potential directions for future investigations.

cs.GT↗

arXiv2007

Exploration via design and the cost of uncertainty in keyword auctions

We present a deterministic exploration mechanism for sponsored search auctions, which enables the auctioneer to learn the relevance scores of advertisers, and allows advertisers to estimate the true value of clicks generated at the auction site. This exploratory mechanism deviates only minimally from the mechanism being currently used by Google and Yahoo! in the sense that it retains the same pricing rule, similar ranking scheme, as well as, similar mathematical structure of payoffs. In particular, the estimations of the relevance scores and true-values are achieved by providing a chance to lower ranked advertisers to obtain better slots. This allows the search engine to potentially test a new pool of advertisers, and correspondingly, enables new advertisers to estimate the value of clicks/leads generated via the auction. Both these quantities are unknown a priori, and their knowledge is necessary for the auction to operate efficiently. We show that such an exploration policy can be incorporated without any significant loss in revenue for the auctioneer. We compare the revenue of the new mechanism to that of the standard mechanism at their corresponding symmetric Nash equilibria and compute the cost of uncertainty, which is defined as the relative loss in expected revenue per impression. We also bound the loss in efficiency, as well as, in user experience due to exploration, under the same solution concept (i.e. SNE). Thus the proposed exploration mechanism learns the relevance scores while incorporating the incentive constraints from the advertisers who are selfish and are trying to maximize their own profits, and therefore, the exploration is essentially achieved via mechanism design. We also discuss variations of the new mechanism such as truthful implementations.

cs.GT↗

arXiv2007

Diversification in the Internet Economy:The Role of For-Profit Mediators

We investigate market forces that would lead to the emergence of new classes of players in the sponsored search market. We report a 3-fold diversification triggered by two inherent features of the sponsored search market, namely, capacity constraints and collusion-vulnerability of current mechanisms. In the first scenario, we present a comparative study of two models motivated by capacity constraints - one where the additional capacity is provided by for-profit agents, who compete for slots in the original auction, draw traffic, and run their own sub-auctions, and the other, where the additional capacity is provided by the auctioneer herself, by essentially acting as a mediator and running a single combined auction. This study was initiated by us in \cite{SRGR07}, where the mediator-based model was studied. In the present work, we study the auctioneer-based model and show that this model seems inferior to the mediator-based model in terms of revenue or efficiency guarantee due to added capacity. In the second scenario, we initiate a game theoretic study of current sponsored search auctions, involving incentive driven mediators who exploit the fact that these mechanisms are not collusion-resistant. In particular, we show that advertisers can improve their payoffs by using the services of the mediator compared to directly participating in the auction, and that the mediator can also obtain monetary benefit, without violating incentive constraints from the advertisers who do not use its services. We also point out that the auctioneer can not do very much via mechanism design to avoid such for-profit mediation without losing badly in terms of revenue, and therefore, the mediators are likely to prevail.

cs.GT↗

arXiv2007

A Symphony Conducted by Brunet

We introduce BruNet, a general P2P software framework which we use to produce the first implementation of Symphony, a 1-D Kleinberg small-world architecture. Our framework is designed to easily implement and measure different P2P protocols over different transport layers such as TCP or UDP. This paper discusses our implementation of the Symphony network, which allows each node to keep $k \le \log N$ shortcut connections and to route to any other node with a short average delay of $O(\frac{1}{k}\log^2 N)$. %This provides a continuous trade-off between node degree and routing latency. We present experimental results taken from several PlanetLab deployments of size up to 1060 nodes. These succes sful deployments represent some of the largest PlanetLab deployments of P2P overlays found in the literature, and show our implementation's robustness to massive node dynamics in a WAN environment.

cs.DC↗

arXiv2007

For-profit mediators in sponsored search advertising

A mediator is a well-known construct in game theory, and is an entity that plays on behalf of some of the agents who choose to use its services, while the rest of the agents participate in the game directly. We initiate a game theoretic study of sponsored search auctions, such as those used by Google and Yahoo!, involving {\em incentive driven} mediators. We refer to such mediators as {\em for-profit} mediators, so as to distinguish them from mediators introduced in prior work, who have no monetary incentives, and are driven by the altruistic goal of implementing certain desired outcomes. We show that in our model, (i) players/advertisers can improve their payoffs by choosing to use the services of the mediator, compared to directly participating in the auction; (ii) the mediator can obtain monetary benefit by managing the advertising burden of its group of advertisers; and (iii) the payoffs of the mediator and the advertisers it plays for are compatible with the incentive constraints from the advertisers who do dot use its services. A simple intuition behind the above result comes from the observation that the mediator has more information about and more control over the bid profile than any individual advertiser, allowing her to reduce the payments made to the auctioneer, while still maintaining incentive constraints. Further, our results indicate that there are significant opportunities for diversification in the internet economy and we should expect it to continue to develop richer structure, with room for different types of agents to coexist.

cs.GT↗

arXiv2007

Capacity constraints and the inevitability of mediators in adword auctions

One natural constraint in the sponsored search advertising framework arises from the fact that there is a limit on the number of available slots, especially for the popular keywords, and as a result, a significant pool of advertisers are left out. We study the emergence of diversification in the adword market triggered by such capacity constraints in the sense that new market mechanisms, as well as, new for-profit agents are likely to emerge to combat or to make profit from the opportunities created by shortages in ad-space inventory. We propose a model where the additional capacity is provided by for-profit agents (or, mediators), who compete for slots in the original auction, draw traffic, and run their own sub-auctions. The quality of the additional capacity provided by a mediator is measured by its {\it fitness} factor. We compute revenues and payoffs for all the different parties at a {\it symmetric Nash equilibrium} (SNE) when the mediator-based model is operated by a mechanism currently being used by Google and Yahoo!, and then compare these numbers with those obtained at a corresponding SNE for the same mechanism, but without any mediators involved in the auctions. Such calculations allow us to determine the value of the additional capacity. Our results show that the revenue of the auctioneer, as well as the social value (i.e. efficiency), always increase when mediators are involved; moreover even the payoffs of {\em all} the bidders will increase if the mediator has a high enough fitness. Thus, our analysis indicates that there are significant opportunities for diversification in the internet economy and we should expect it to continue to develop richer structure, with room for different types of agents and mechanisms to coexist.

cs.GT↗

arXiv2006

Comparison of Image Similarity Queries in P2P Systems

Given some of the recent advances in Distributed Hash Table (DHT) based Peer-To-Peer (P2P) systems we ask the following questions: Are there applications where unstructured queries are still necessary (i.e., the underlying queries do not efficiently map onto any structured framework), and are there unstructured P2P systems that can deliver the high bandwidth and computing performance necessary to support such applications. Toward this end, we consider an image search application which supports queries based on image similarity metrics, such as color histogram intersection, and discuss why in this setting, standard DHT approaches are not directly applicable. We then study the feasibility of implementing such an image search system on two different unstructured P2P systems: power-law topology with percolation search, and an optimized super-node topology using structured broadcasts. We examine the average and maximum values for node bandwidth, storage and processing requirements in the percolation and super-node models, and show that current high-end computers and high-speed links have sufficient resources to enable deployments of large-scale complex image search systems.

cs.DC↗

arXiv2006

Low Latency Wireless Ad-Hoc Networking: Power and Bandwidth Challenges and a Hierarchical Solution

This paper is concerned with the scaling of the number of hops in a large scale wireless ad-hoc network (WANET), a quantity we call network latency. A large network latency affects all aspects of data communication in a WANET, including an increase in delay, packet loss, required processing power and memory. We consider network management and data routing challenges in WANETs with scalable network latency. On the physical side, reducing network latency imposes a significantly higher power and bandwidth demand on nodes, as is reflected in a set of new bounds. On the protocol front, designing distributed routing protocols that can guarantee the delivery of data packets within scalable number of hops is a challenging task. To solve this, we introduce multi-resolution randomized hierarchy (MRRH), a novel power and bandwidth efficient WANET protocol with scalable network latency. MRRH uses a randomized algorithm for building and maintaining a random hierarchical network topology, which together with the proposed routing algorithm can guarantee efficient delivery of data packets in the wireless network. For a network of size $N$, MRRH can provide an average latency of only $O(\log^{3} N)$. The power and bandwidth consumption of MRRH are shown to be \emph{nearly} optimal for the latency it provides. Therefore, MRRH, is a provably efficient candidate for truly large scale wireless ad-hoc networking.

cs.IT↗

arXiv2006

A General Framework for Scalability and Performance Analysis of DHT Routing Systems

In recent years, many DHT-based P2P systems have been proposed, analyzed, and certain deployments have reached a global scale with nearly one million nodes. One is thus faced with the question of which particular DHT system to choose, and whether some are inherently more robust and scalable. Toward developing such a comparative framework, we present the reachable component method (RCM) for analyzing the performance of different DHT routing systems subject to random failures. We apply RCM to five DHT systems and obtain analytical expressions that characterize their routability as a continuous function of system size and node failure probability. An important consequence is that in the large-network limit, the routability of certain DHT systems go to zero for any non-zero probability of node failure. These DHT routing algorithms are therefore unscalable, while some others, including Kademlia, which powers the popular eDonkey P2P system, are found to be scalable.

cs.DC↗

arXiv2006

Finite Percolation at a Multiple of the Threshold

Bond percolation on infinite heavy-tailed power-law random networks lacks a proper phase transition; or one may say, there is a phase transition at {\em zero percolation probability}. Nevertheless, a finite size percolation threshold $q_c(N)$, where $N$ is the network size, can be defined. For such heavy-tailed networks, one can choose a percolation probability $q(N)=ρq_c(N)$ such that $\displaystyle \lim_{N\to \infty}(q-q_c(N)) =0$, and yet $ρ$ is arbitrarily large (such a scenario does not exist for networks with non-zero percolation threshold). We find that the critical behavior of random power-law networks is best described in terms of $ρ$ as the order parameter, rather than $q$. This paper makes the notion of the phase transition of the size of the largest connected component at $ρ=1$ precise. In particular, using a generating function based approach, we show that for $ρ>1$, and the power-law exponent, $2\leq τ<3$, the largest connected component scales as $\sim N^{1-1/τ}$, while for $0<ρ<1$ the scaling is at most $\sim N^{\frac{2-τ}τ}$; here, the maximum degree of any node, $k_{max}$, has been assumed to scale as N^{1/τ}$. In general, our approach yields that for large $N$, $ρ\gg 1$, $2\leq τ<3$, and $k_{max} \sim N^{1/τ}$, the largest connected component scales as $\sim ρ^{1/(3-τ)}N^{1-1/τ}$.Thus, for any fixed but large N, we recover, and make it precise, a recent result that computed a scaling behavior of $q^{1/(3-τ)}$ for "small $q$". We also provide large-scale simulation results validating some of these scaling predictions, and discuss applications of these scaling results to supporting efficient unstructured queries in peer-to-peer networks.

cond-mat.dis-nn↗

arXiv2006

Balanced Overlay Networks (BON): Decentralized Load Balancing via Self-Organized Random Networks

We present a novel framework, called balanced overlay networks (BON), that provides scalable, decentralized load balancing for distributed computing using large-scale pools of heterogeneous computers. Fundamentally, BON encodes the information about each node's available computational resources in the structure of the links connecting the nodes in the network. This distributed encoding is self-organized, with each node managing its in-degree and local connectivity via random-walk sampling. Assignment of incoming jobs to nodes with the most free resources is also accomplished by sampling the nodes via short random walks. Extensive simulations show that the resulting highly dynamic and self-organized graph structure can efficiently balance computational load throughout large-scale networks. These simulations cover a wide spectrum of cases, including significant heterogeneity in available computing resources and high burstiness in incoming load. We provide analytical results that prove BON's scalability for truly large-scale networks: in particular we show that under certain ideal conditions, the network structure converges to Erdos-Renyi (ER) random graphs; our simulation results, however, show that the algorithm does much better, and the structures seem to approach the ideal case of d-regular random graphs. We also make a connection between highly-loaded BONs and the well-known ball-bin randomized load balancing framework.

cs.DC↗