Rashmi Sundareswara

Perceptual multistability predicted by search model for Bayesian decisions.

Perceptual multistability refers to the phenomenon of spontaneous perceptual switching between two or more likely interpretations of an image. Although frequently explained by processes of adaptation or hysteresis, we show that perceptual switching can arise as a natural byproduct of perceptual decision making based on probabilistic (Bayesian) inference, which interprets images by combining probabilistic models of image formation with knowledge of scene regularities. Empirically, we investigated the effect of introducing scene regularities on Necker cube bistability by flanking the Necker cube with fields of unambiguous cubes that are oriented to coincide with one of the Necker cube percepts. We show that background cubes increase the time spent in percepts most similar to the background. To characterize changes in the temporal dynamics of the perceptual alternations beyond percept durations, we introduce Markov Renewal Processes (MRPs). MRPs provide a general mathematical framework for describing probabilistic switching behavior in finite state processes. Additionally, we introduce a simple theoretical model consistent with Bayesian models of vision that involves searching for good interpretations of an image by sampling a posterior distribution coupled with a decay process that favors recent to old interpretations. The model has the same quantitative characteristics as our human data and variation in model parameters can capture between-subject variation. Because the model produces the same kind of stochastic process found in human perceptual behavior, we conclude that multistability may represent an unavoidable by-product of normal perceptual (Bayesian) decision making with ambiguous images.

Accurate statistical approaches for generating representative workload compositions

Composing a representative workload is a crucial step during the design process of a microprocessor. The workload should be composed in such a way that it is representative for the target domain of application and yet, the amount of redundancy in the workload should be minimized as much as possible in order not to overly increase the total simulation time. As a result, there is an important trade-off that needs to be made between workload representativeness and simulation accuracy versus simulation speed. Previous work used statistical data analysis techniques to identify representative benchmarks and corresponding inputs, also called a subset, from a large set of potential benchmarks and inputs. These methodologies measure a number of program characteristics on which principal components analysis (PCA) is applied before identifying distinct program behaviors among the benchmarks using cluster analysis. In this paper we propose independent components analysis (ICA) as a better alternative to PCA as it does not assume that the original data set has a Gaussian distribution, which allows ICA to better find the important axes in the workload space. Our experimental results using SPEC CPU2000 benchmarks show that ICA significantly outperforms PCA in that ICA achieves smaller benchmark subsets that are more accurate than those found by PCA.

Bid Evaluation and Selection in the MAGNET Automated Contracting System

We present an approach to the bid-evaluation problem in a system for multi-agent contract negotiation, called MAGNET. The MAGNET market infrastructure provides support for a variety of types of transactions, from simple buying and selling of goods and services to complex multi-agent contract negotiations. In the latter case, MAGNET is designed to negotiate contracts based on temporal and precedence constraints, and includes facilities for dealing with time-based contingencies. One responsibility of a customer agent in the MAGNET system is to select an optimal bid combination. We present an efficient anytime algorithm for a customer agent to select bids submitted by supplier agents in response to a call for bids. Bids might include combinations of subtasks and might include discounts for combinations. In an experimental study we explore the behavior of the algorithm based on the interactions of factors such as bid prices, number of bids, and number of subtasks. The results of experiments we present show that the algorithm is extremely efficient even for large number of bids.

Evaluating risk: flexibility and feasibility in multi-agent contracting

In an automated contracting environment, where a contractor agent must negotiate with other self-interested supplier agents in order to execute its plans, there is a tradeoff between giving the suppliers sufficient flexibility to incorporate the requirements of the contractor’s call-for-bids into their own resource schedules, and ensuring the contractor that any bids received can be composed into a feasible plan. We introduce a bid evaluation process that incorporates cost, task coverage, temporal feasibility, and risk estimation. Using this evaluation process, we describe an empirical study of the tradeoffs between flexibility, plan feasibility, and cost in the context of our MAGNET multi-agent contracting market infrastructure.

Bid selection strategies for multi-agent contracting in the presence of scheduling constraints

Bid evaluation in a multi-agent automated contracting environment presents a challenging search problem. We introduce a multicriterion, anytime bid evaluation strategy that incorporates cost, task coverage, temporal feasibility, and risk estimation into a simulated annealing framework. We report on an experimental evaluation using a set of increasingly informed search heuristics within simulated annealing. The results show that excess focus on improvement leads to faster improvement early on, at the cost of a lower likelihood of finding a solution that satisfies all the constraints. The most successful approach used a combination of random and focused bid selection methods, along with pruning and repeated restarts.

Bayesian modelling of camera calibration and reconstruction

Camera calibration methods, whether implicit or explicit, are a critical part of most 3D vision systems. These methods involve estimation of a model for the camera that produced the visual input, and subsequently to infer the 3D structure that gave rise to the input. However, in these systems the error in calibration is typically unknown, or if known, the effect of calibration error on subsequent processing (e.g. 3D reconstruction) is not accounted for. In this paper, we propose a Bayesian camera calibration method that explicitly computes calibration error, and we show how knowledge of this error can be used to improve the accuracy of subsequent processing. What distinguishes the work is the explicit computation of a posterior distribution on unknown camera parameters, rather than just a best estimate. Marginalizing (averaging) subsequent estimates by this posterior is shown to reduce reconstruction error over calibration approaches that rely on a single best estimate. The method is made practical using sampling techniques, that require only the evaluation of the calibration error function and the specification of priors. Samples with their corresponding probability weights can be used to produce better estimates of the camera parameters. Moreover, these samples can be directly used to improve estimates that rely on calibration information, like 3D reconstruction. We evaluate our method using simulated data for a structure from motion problem, in which the same point matches are used to calibrate the camera, estimate the motion, and reconstruct the 3D geometry. Our results show improved reconstruction over non-linear Camera calibration methods like the Maximum Likelihood estimate. Additionally, this approach scales much better in the face of increasingly noisy point matches.

A bio-inspired kinematic controller for obstacle avoidance during reaching tasks with real robots

This paper describes a redundant robot arm that is capable of learning to reach for targets in space while avoiding obstacles in a self-organized fashion. Self-generated movement commands that activate correlated visual, spatial and motor information are used to learn forward and inverse kinematic control models while moving in obstacle free space using the direction-to-rotation transform (DIRECT). The DIRECT based kinematic controller is fault tolerant and can handle a wide range of perturbations such as joint locking and the use of tools despite not experiencing them during learning. We have developed a DIRECT-based reactive obstacle avoidance controller (DIRECT-ROAC) that enables the redundant robot arm to avoid obstacles in environments with simple obstacle configurations. However, certain configurations of obstacles in the environment prevent the robot from reaching the target with purely reactive obstacle avoidance. To address this complexity, we model a self-organized process of mental rehearsals of movements inspired by human and animal experiments on reaching to generate plans for movement execution using DIRECT-ROAC in complex environments. These mental rehearsals or plans are self generated by utilizing perceptual information in the form of via-points extracted from attentional shrouds around obstacles in its environment. Computer simulations show that the proposed novel controller is successful in avoiding obstacles in environments with complex obstacle configurations.

High Precision Object Segmentation and Tracking for use in Super Resolution Video Reconstruction

Super resolution image reconstruction allows for the enhancement of images in a video sequence that is superior to the original pixel resolution of the imager. Difficulty arises when there are foreground objects that move differently than the background. A common example of this is a car in motion in a video. Given the common occurrence of such situations, super resolution reconstruction becomes non-trivial. One method for dealing with this is to segment out foreground objects and quantify their pixel motion differently. First we estimate local pixel motion using a standard block motion algorithm common to MPEG encoding. This is then combined with the image itself into a five dimensional mean-shift kernel density estimation based image segmentation with mixed motion and color image feature information. This results in a tight segmentation of objects in terms of both motion and visible image features. The next step is to combine segments into a single master object. Statistically common motion and proximity are used to merge segments into master objects. To account for inconsistencies that can arise when tracking objects, we compute statistics over the object and fit it with a generalized linear model. Using the Kullback-Leibler divergence, we have a metric for the goodness of the track for an object between frames.

Bayesian discounting of camera parameter uncertainty for optimal 3D reconstruction from images

3D reconstruction through point correspondences is a process that is sensitive to match errors and also to possible ambiguity in the solution space of shape and camera estimates - the existence of either or the combination of both propagates into sub-optimal estimates of the structure. To counteract this, most methods in the field jointly or sequentially estimate both the camera parameters and the 3D structure using methods such as Bundle Adjustment. However, joint estimation methods such as Bundle Adjustment find sub-optimal solutions of structure if the structure is not uniquely defined in the joint space. Using probabilistic models for reconstruction and marginalizing across camera parameter uncertainty we show how to compute the optimal 3D reconstruction. We use only uniform priors to make comparisons between Bundle Adjustment and our work. However, the method, by its construction, is set up to use prior information about either the camera parameters or the 3D structure, if it is available. Results show that this method produces better reconstruction estimates than joint estimation methods such as Bundle Adjustment especially in the face of increasing noise in the feature correspondences.

Network-based relevant feature identification and early detection of impending faults

As more electronic sensors for control units are added into next generation vehicles, and linked to communication infrastructure, greater amount of operational data has become unprecedentedly available close to real time. The ability to extract patterns in such real-time massive operational data to detect, isolate, predict, and mitigate faults is the key to enhance vehicle ownership experiences. In this paper, we present a Network Predictive Analytics (NPA) methodology for individualized, evidence-based, preventive vehicle health management. We leverage a recent novel measure called the Maximal Information Coefficient (MIC) - a method that computes an information-theoretic score for quantifying the association strength between two time-series, regardless of the type of relationship. Built upon MIC, we develop an automated method to extract fault-specific relevant features from Association Feature Networks by clustering and ranking dynamic feature relations. By monitoring identified features, our novel rate-based spectral Early Warning Signal (EWS-rate) signals further detects impending faults prior to their repairs. We study the effectiveness of our methodology over a set of data collected from real vehicles. The result shows the proposed Network Predictive Analytics (NPA) methodology can provide significant lead time in early detection of impending faults with high accuracy.