arXiv subjects

Abhishek Sinha

Publications and source records attributed to Abhishek Sinha.

66 records · Page 4Linked to original sources

arXiv2017

Introspection: Accelerating Neural Network Training By Learning Weight Evolution

Neural Networks are function approximators that have achieved state-of-the-art accuracy in numerous machine learning tasks. In spite of their great success in terms of accuracy, their large training time makes it difficult to use them for various tasks. In this paper, we explore the idea of learning weight evolution pattern from a simple network for accelerating training of novel neural networks. We use a neural network to learn the training pattern from MNIST classification and utilize it to accelerate training of neural networks used for CIFAR-10 and ImageNet classification. Our method has a low memory footprint and is computationally efficient. This method can also be used with other optimizers to give faster convergence. The results indicate a general trend in the weight evolution during training of neural networks.

cs.LG↗

arXiv2017

Throughput-Optimal Broadcast in Wireless Networks with Point-to-Multipoint Transmissions

We consider the problem of efficient packet dissemination in wireless networks with point-to-multi-point wireless broadcast channels. We propose a dynamic policy, which achieves the broadcast capacity of the network. This policy is obtained by first transforming the original multi-hop network into a precedence-relaxed virtual single-hop network and then finding an optimal broadcast policy for the relaxed network. The resulting policy is shown to be throughput-optimal for the original wireless network using a sample-path argument. We also prove the NP-completeness of the finite-horizon broadcast problem, which is in contrast with the polynomial time solvability of the problem with point-to-point channels. Illustrative simulation results demonstrate the efficacy of the proposed broadcast policy in achieving the full broadcast capacity with low delay.

cs.NI↗

arXiv2016

Optimal Control for Generalized Network-Flow Problems

We consider the problem of throughput-optimal packet dissemination, in the presence of an arbitrary mix of unicast, broadcast, multicast and anycast traffic, in a general wireless network. We propose an online dynamic policy, called Universal Max-Weight (UMW), which solves the above problem efficiently. To the best of our knowledge, UMW is the first throughput-optimal algorithm of such versatility in the context of generalized network flow problems. Conceptually, the UMW policy is derived by relaxing the precedence constraints associated with multi-hop routing, and then solving a min-cost routing and max-weight scheduling problem on a virtual network of queues. When specialized to the unicast setting, the UMW policy yields a throughput-optimal cycle-free routing and link scheduling policy. This is in contrast to the well-known throughput-optimal Back- Pressure (BP) policy which allows for packet cycling, resulting in excessive latency. Extensive simulation results show that the proposed UMW policy incurs a substantially smaller delay as compared to the BP policy. The proof of throughput-optimality of the UMW policy combines ideas from stochastic Lyapunov theory with a sample path argument from adversarial queueing theory and may be of independent theoretical interest.

math.OC↗

arXiv2016

Throughput-Optimal Broadcast in Wireless Networks with Dynamic Topology

We consider the problem of throughput-optimal broadcast- ing in time-varying wireless networks, whose underlying topology is restricted to Directed Acyclic Graphs (DAG). Previous broadcast algorithms route packets along spanning trees. In large networks with time-varying connectivities, these trees are difficult to compute and maintain. In this paper we propose a new online throughput-optimal broadcast algorithm which makes packet-by-packet scheduling and routing decisions, obviating the need for maintaining any global topological structures, such as spanning-trees. Our algorithm relies on system-state information for making transmission decisions and hence, may be thought of as a generalization of the well-known back-pressure algorithm which makes point-to-point unicast transmission decisions based on queue-length information, without requiring knowledge of end-to-end paths. Technically, the back-pressure algorithm is derived by stochastically stabilizing the network-queues. However, because of packet-duplications associated with broadcast, the work-conservation principle is violated and queuing processes are difficult to define in the broadcast problem. To address this fundamental issue, we identify certain state-variables which behave like virtual queues in the broadcast setting. By stochastically stabilizing these virtual queues, we devise a throughput-optimal broadcast policy. We also derive new characterizations of the broadcast-capacity of time-varying wireless DAGs and derive an efficient algorithm to compute the capacity exactly under certain assumptions of interference model, and a poly-time approximation algorithm for computing the capacity under less restrictive assumptions.

cs.IT↗

arXiv2016

Throughput-Optimal Multi-hop Broadcast Algorithms

In this paper we design throughput-optimal dynamic broad- cast algorithms for multi-hop networks with arbitrary topolo- gies. Most of the previous broadcast algorithms route pack- ets along spanning trees, rooted at the source node. For large dynamic networks, computing and maintaining a set of spanning trees is not efficient, as the network-topology may change frequently. In this paper we design a class of dynamic algorithms which makes packet-by-packet schedul- ing and routing decisions and thus obviates the need for maintaining any global topological structures, such as span- ning trees. Our algorithms may be conveniently understood as a non-trivial generalization of the familiar back-pressure algorithm which makes unicast packet routing and schedul- ing decisions, based on queue-length information, without maintaining end-to-end paths. However, in the broadcast problem, it is hard to define queuing structures due to ab- sence of a work-conservation principle which results from packet duplications. We design and prove the optimality of a virtual-queue based algorithm, where a virtual-queue is de- fined for subsets of vertices. We then propose a multi-class broadcast policy which combines the above scheduling algo- rithm with a class-based in-order packet delivery constraint, resulting in significant reduction in complexity. Finally, we evaluate performance of the proposed algorithms via exten- sive numerical simulations.

cs.DS↗

arXiv2016

Distributed Load Management Algorithms in Anycast-based CDNs

Anycast is an internet addressing protocol where multiple hosts share the same IP-address. A popular architecture for modern Content Distribution Networks (CDNs) for geo-replicated services consists of multiple layers of proxy nodes for service and co-located DNS-servers for load-balancing among different proxies. Both the proxies and the DNS-servers use anycast addressing, which offers simplicity of design and high availability of service at the cost of partial loss of routing control. Due to the very nature of anycast, redirection actions by a DNS-server also affects loads at nearby proxies in the network. This makes the problem of optimal distributed load management highly challenging. In this paper, we propose and evaluate an analytical framework to formulate and solve the load-management problem in this context. We consider two distinct algorithms. In the first half of the paper, we pose the load-management problem as a convex optimization problem. Following a Kelly-type dual decomposition technique, we propose a fully-distributed load-management algorithm by introducing FastControl packets. This algorithm utilizes the underlying anycast mechanism itself to enable effective coordination among the nodes, thus obviating the need for any external control channel. In the second half of the paper, we consider an alternative greedy load-management heuristic, currently in production in a major commercial CDN. We study its dynamical characteristics and analytically identify its operational and stability properties. Finally, we critically evaluate both the algorithms and explore their optimality-vs-complexity trade-off using trace-driven simulations.

cs.NI↗

arXiv2015

Distributed Load Management in Anycast-based CDNs

Anycast is an internet addressing protocol where multiple hosts share the same IP-address. A popular architecture for modern Content Distribution Networks (CDNs) for geo-replicated HTTP-services consists of multiple layers of proxy nodes for service and co-located DNS-servers for load-balancing on different proxies. Both the proxies and the DNS-servers use anycast addressing, which offers simplicity of design and high availability of service at the cost of partial loss of routing control. Due to the very nature of anycast, load-management decisions by a co-located DNS-server also affects loads at nearby proxies in the network. This makes the problem of distributed load management highly challenging. In this paper, we propose an analytical framework to formulate and solve the load-management problem in this context. We consider two distinct algorithms. In the first half of the paper, we pose the load-management problem as a convex optimization problem. Following a dual decomposition technique, we propose a fully-distributed load management algorithm by introducing FastControl packets. This algorithm utilizes the underlying anycast mechanism itself to enable effective coordination among the nodes, thus obviating the need for any external control channel. In the second half of the paper, we consider an alternative greedy load-management heuristic, currently in production in a major commercial CDN. We study its dynamical characteristics and analytically identify its operational and stability properties. Finally, we critically evaluate both the algorithms and explore their optimality-vs-complexity trade-off using trace-driven simulations.

cs.NI↗

arXiv2015

Throughput-Optimal Multihop Broadcast on Directed Acyclic Wireless Networks

We study the problem of efficiently broadcasting packets in multi-hop wireless networks. At each time slot the network controller activates a set of non-interfering links and forwards selected copies of packets on each activated link. A packet is considered jointly received only when all nodes in the network have obtained a copy of it. The maximum rate of jointly received packets is referred to as the broadcast capacity of the network. Existing policies achieve the broadcast capacity by balancing traffic over a set of spanning trees, which are difficult to maintain in a large and time-varying wireless network. We propose a new dynamic algorithm that achieves the broadcast capacity when the underlying network topology is a directed acyclic graph (DAG). This algorithm is decentralized, utilizes local queue-length information only and does not require the use of global topological structures such as spanning trees. The principal technical challenge inherent in the problem is the absence of work-conservation principle due to the duplication of packets, which renders traditional queuing modelling inapplicable. We overcome this difficulty by studying relative packet deficits and imposing in-order delivery constraints to every node in the network. Although in-order packet delivery, in general, leads to degraded throughput in graphs with cycles, we show that it is throughput optimal in DAGs and can be exploited to simplify the design and analysis of optimal algorithms. Our characterization leads to a polynomial time algorithm for computing the broadcast capacity of any wireless DAG under the primary interference constraints. Additionally, we propose an extension of our algorithm which can be effectively used for broadcasting in any network with arbitrary topology.

cs.IT↗

arXiv2014

Throughput-Optimal Broadcast on Directed Acyclic Graphs

We study the problem of broadcasting packets in wireless networks. At each time slot, a network controller activates non-interfering links and forwards packets to all nodes at a common rate; the maximum rate is referred to as the broadcast capacity of the wireless network. Existing policies achieve the broadcast capacity by balancing traffic over a set of spanning trees, which are difficult to maintain in a large and time-varying wireless network. We propose a new dynamic algorithm that achieves the broadcast capacity when the underlying network topology is a directed acyclic graph (DAG). This algorithm utilizes local queue-length information, does not use any global topological structures such as spanning trees, and uses the idea of in-order packet delivery to all network nodes. Although the in-order packet delivery constraint leads to degraded throughput in cyclic graphs, we show that it is throughput optimal in DAGs and can be exploited to simplify the design and analysis of optimal algorithms. Our simulation results show that the proposed algorithm has superior delay performance as compared to tree-based approaches.

cs.NI↗

arXiv2013

Optimal Capacity Relay Node Placement in a Multi-hop Wireless Network on a Line

We use information theoretic achievable rate formulas for the multi-relay channel to study the problem of optimal placement of relay nodes along the straight line joining a source node and a sink node. The achievable rate formulas that we use are for full-duplex radios at the relays and decode- and-forward relaying. For the single relay case, and individual power constraints at the source node and the relay node, we provide explicit formulas for the optimal relay location and the optimal power allocation to the source-relay channel, for the exponential and the power-law path-loss channel models. For the multiple relay case, we consider exponential path-loss and a total power constraint over the source and the relays, and derive an optimization problem, the solution of which provides the optimal relay locations. Numerical results suggest that at low attenuation the relays are mostly clustered close to the source in order to be able to cooperate among themselves, whereas at high attenuation they are uniformly placed and work as repeaters. The structure of the optimal power allocation for a given placement of the nodes, then motivates us to formulate the problem of impromptu ("as-you-go") placement of relays along a line of exponentially distributed length, with exponential path- loss, so as to minimize a cost function that is additive over hops. The hop cost trades off a capacity limiting term, motivated from the optimal power allocation solution, against the cost of adding a relay node. We formulate the problem as a total cost Markov decision process, for which we prove results for the value function, and provide insights into the placement policy via numerical exploration.

cs.NI↗

arXiv2012

Optimal Sequential Wireless Relay Placement on a Random Lattice Path

Our work is motivated by the need for impromptu (or "as-you-go") deployment of relay nodes (for establishing a packet communication path with a control centre) by fire-men/commandos while operating in an unknown environment. We consider a model, where a deployment operative steps along a random lattice path whose evolution is Markov. At each step, the path can randomly either continue in the same direction or take a turn "North" or "East," or come to an end, at which point a data source (e.g., a temperature sensor) has to be placed that will send packets to a control centre at the origin of the path. A decision has to be made at each step whether or not to place a wireless relay node. Assuming that the packet generation rate by the source is very low, and simple link-by-link scheduling, we consider the problem of relay placement so as to minimize the expectation of an end-to-end cost metric (a linear combination of the sum of convex hop costs and the number of relays placed). This impromptu relay placement problem is formulated as a total cost Markov decision process. First, we derive the optimal policy in terms of an optimal placement set and show that this set is characterized by a boundary beyond which it is optimal to place. Next, based on a simpler alternative one-step-look-ahead characterization of the optimal policy, we propose an algorithm which is proved to converge to the optimal placement set in a finite number of steps and which is faster than the traditional value iteration. We show by simulations that the distance based heuristic, usually assumed in the literature, is close to the optimal provided that the threshold distance is carefully chosen.

cs.NI↗

arXiv2011

On Periodic Node Deployment in Wireless Sensor Networks: A Statistical Analysis

Rapid progress made in the field of sensor technology, wireless communication, and computer networks in recent past, led to the development of wireless Ad-hoc sensor networks, consisting of small, low-cost sensors, which can monitor wide and remote areas with precision and liveliness unseen to the date without the intervention of a human operator. This work comes up with a stochastic model for periodic sensor-deployment (in face of their limited amount of battery-life) to maintain a minimal node-connectivity in wireless sensor networks. The node deployment cannot be modeled by using results from conventional continuous birth-death process, since new nodes are added to the network in bursts, i.e. the birth process is not continuous in practical situations. We analyze the periodic node deployment process using discrete birth-continuous death process and obtain two important statistical measures of the existing number of nodes in the network, namely the mean and variance. We show that the above mentioned sequences of mean and variances always converge to finite steady state values, thus ensuring the stability of the system. We also develop a cost function for the process of periodic deployment of sensor nodes and minimize it to find the optimal time (τ) and optimum number of re-deployment (q) for maintaining minimum connectivity in the network.

cs.NI↗