Integrated Gradients (IG) [29] is a commonly used feature attribution method for deep neural networks. While IG has many desirable properties, the method often produces spurious/noisy pixel attributions in regions that are not related to the predicted class when applied to visual models. While this has been previously noted [27], most existing solutions [25], [17] are aimed at addressing the symptoms by explicitly reducing the noise in the resulting attributions. In this work, we show that one of the causes of the problem is the accumulation of noise along the IG path. To minimize the effect of this source of noise, we propose adapting the attribution path itself - conditioning the path not just on the image but also on the model being explained. We introduce Adaptive Path Methods (APMs) as a generalization of path methods, and Guided IG as a specific instance of an APM. Empirically, Guided IG creates saliency maps better aligned with the model's prediction and the input image that is being explained. We show through qualitative and quantitative experiments that Guided IG outperforms other, related methods in nearly every experiment.

When a deep learning model is deployed in the wild, it can encounter test data drawn from distributions different from the training data distribution and suffer drop in performance. For safe deployment, it is essential to estimate the accuracy of the pre-trained model on the test data. However, the labels for the test inputs are usually not immediately available in practice, and obtaining them can be expensive. This observation leads to two challenging tasks: (1) unsupervised accuracy estimation, which aims to estimate the accuracy of a pre-trained classifier on a set of unlabeled test inputs; (2) error detection, which aims to identify mis-classified test inputs. In this paper, we propose a principled and practically effective framework that simultaneously addresses the two tasks. The proposed framework iteratively learns an ensemble of models to identify mis-classified data points and performs self-training to improve the ensemble with the identified points. Theoretical analysis demonstrates that our framework enjoys provable guarantees for both accuracy estimation and error detection under mild conditions readily satisfied by practical deep learning models. Along with the framework, we proposed and experimented with two instantiations and achieved state-of-the-art results on 59 tasks. For example, on iWildCam, one instantiation reduces the estimation error for unsupervised accuracy estimation by at least 70% and improves the F1 score for error detection by at least 4.7% compared to existing methods.

Feature attribution methods, proposed recently, help users interpret the predictions of complex models.Our approach integrates feature attributions into the objective function to allow machine learning practitioners to incorporate priors in model building.To demonstrate the effectiveness our technique, we apply it to two tasks: (1) mitigating unintended bias in text classifiers by neutralizing identity terms; (2) improving classifier performance in a scarce data setting by forcing the model to focus on toxic terms.Our approach adds an L 2 distance loss between feature attributions and task-specific prior values to the objective.Our experiments show that i) a classifier trained with our technique reduces undesired model biases without a tradeoff on the original task; ii) incorporating priors helps model performance in scarce data settings.

This work presents a system for augmenting the functionality of Yelp-like recommendation sites by enabling users to search for places bounded by travel-time when using public transportation, and modifying recommendations based on updated weather conditions. Using public transport, although is cheaper and efficient, entails that only fixed places of boarding/exiting may be used which, in turn, implies walking to (from) a particular location from (to) a given station. Given the impact of the weather on the mood and activities, preferences for a certain type of services may need to be dynamically adjusted based on the current weather or the near-future forecast, modulo travel-routes to preferred locations. In this work, we develop a model to predict a user's preferred mode of transport (car, or public transit) from their old check-ins and incorporate the weather context into the recommendation process. We use event-based modeling to control the extent of walking depending on user-defined tolerance information and live weather conditions. We implemented a web application (both desktop and mobile platforms), utilizing existing tools such as Google Maps Direction API and Open Weather Map API for retrieving real-time information.

We present an implementation of a system for managing evolving shapes in Wireless Sensor Networks (WSN). A shape is a contiguous region in which the measurements of the sensors detect values above a given threshold. Our system, in its current version, solves two important problems: (1) Detecting and tracking the changes of boundaries; (2) Detecting an occurrence of within distance predicate for two (or more) shapes. A centralized approach (transmitting raw measurements to a dedicated sink) incurs communication overhead, so we developed distributed algorithms for managing the predicates related to evolving shapes. This demo will present the implementation of our solutions in a heterogeneous WSN consisting of TelosB and SunSPOT motes. It will also illustrate the end-user tools: interface for specifying the parameters of the predicates, along with real-time visualization of their evaluation.

This work addresses the problem of efficient distributed detection and tracking of mobile and evolving/deformable spatial shapes in Wireless Sensor Networks (WSN). The shapes correspond to contiguous regions bounding the locations of sensors in which the readings of the sensors satisfy a particular threshold-based criterion related to the values of a physical phenomenon that they measure. We formalize the predicates representing the shapes in such settings and present detection algorithms. In addition, we provide a light-weight protocol and aggregation methods for energy-efficient distributed execution of those algorithms. Another contribution of this work is that we developed efficient techniques for detecting a co-occurrence of shapes within a given proximity from each other. Our experiments demonstrate that, when compared to the centralized techniques -- which is, predicates being detected in a dedicated sink -- as well as distributed periodic contours construction, our methodologies yield significant energy/communication savings.

We tackle two important problems that arise in simulation-based studies of various data-related properties in the context of Wireless Sensor Networks (WSNs): (1) reducing the turnaround time for completing the simulations in a large-scale parameter space; (2) providing database functionalities for a more detailed insight into the simulation's evolution. Towards these goals, we have developed the DiSSIDnet (Distributed System for Simulation and Integrated Development for Wireless Sensor Networks). Leveraging upon our earlier works on the SIDnet-SWANS tool [3], DiSSIDnet not only provides the ability of a synchronized execution of the simulations in a distributed environment, but also maintains the simulation data over the parameter-space in a relational database. In addition to post-simulation queries that can be posed to the database, we also provide the feature of specifying triggers that can generate notifications upon detecting certain events of interest during the simulation process.

We address the problem of efficient detection of destination-related motion trends in Wireless Sensor Networks (WSN) where tracking is done in collaborative manner among the sensor nodes participating in location detection. In addition to determining a single location, applications may need to detect whether certain properties are true for the (portion of the) entire trajectories. Transmitting the sequence of (location, time) values to a dedicated sink and relying on the sink to detect the validity of the desired properties is a brute-force approach that generates a lot of communication overhead. We present an in-network distributed algorithm for efficient detecting of the Continuously Moving Towards predicate with respect to a given destination that is either a point or a region with polygonal boundary. Our experiments demonstrate that the proposed approaches yield substantial savings when compared to the brute-force one.

This work addresses the problem of geographic routing in the presence of holes or voids in wireless sensor networks. We postulate that, once the boundary of the hole has been established, relying on the existing algorithms for bypassing it may cause severe depletion of the energy reserves among the nodes at (or near) that boundary. This, in turn, may soon render some of those nodes useless for any routing (and/or sensing) purposes, thereby effectively enlarging the size of the pre-existing hole. To extend the lifetime of the nodes along the boundary of a given hole, we propose two heuristic approaches which aim at relieving some of the routing load of the boundary nodes. Towards that, our approaches propose that some of the routes that would otherwise need to bypass the hole along the boundary, should instead start to deviate from their original path further from the hole. Our experiments demonstrate that the proposed approaches not only increase the lifetime of the nodes along the boundary of a given hole, but also yield a more uniform depletion of the energy reserves in its vicinity.

This work addresses the problem of managing the sensor-coverage and organizing the epochs in a manner that balances the trades-offs between the accuracy and energy consumptions during target tracking in Wireless Sensor Networks (WSN). While the typical target tracking approaches are based on movement prediction, we only assume a knowledge of some maximal speed of the target during certain time-intervals. This, in turn, restricts its whereabouts to a disk-bound area throughout such intervals. In such settings, we seek to determine a sensor cover, a subset of all the nodes that need to be awake, which ensures that the target can be detected during the given epoch. Towards this, we propose sensor-cover selection methodologies, Greedy Uncertain Moving Object coverage sensor set selection (GUMO) and PAttern Based coverage sensor set selection (PAB). GUMO is a heuristic maximizing the coverage gain at each step, while PAB is an approach based on optimal deployment pattern of sensor nodes achieving coverage of the disk area bounding the target's whereabouts. We conduct extensive simulations to evaluate the performance of the algorithms, and the results reveal that GUMO and PAB not only provide substantial energy saving due to reduction in the communications involved in selecting tracking participant-nodes and principal(s), while assuring a bounded error on the target's location.