In this work, we introduce the idea of “mutually exclusive learning” and formulate it as a multi-label classification problem. We provide two implementations of the mutually exclusive learner; the first using a consensus based multilabel framework and the second using an adaptive multi-label framework. We use the mutually exclusive learning paradigm for designing both generators and discriminators (classifiers). We experiment with the MNIST dataset for both classifying the digits using a mutually exclusive classifier and generating handwritten digits using a generative adversarial network (GAN) trained with a mutually exclusive discriminator implementation. Our results establish that GANs trained with a mutually exclusive discriminator converge faster than the corresponding GAN with a standard discriminator. Furthermore, the quality of the generated images is also visually better than those generated by a regular GAN.

We introduce a robust neural network based method for classifying aircraft, using raw I/Q data obtained from the Automatic Dependent Surveillance-Broadcast (ADS-B) data from airplanes. ADS-B has become the de-facto standard for air traffic control and forms the basis of the Next Generation Air Transportation System (NextGen). Although ADS-B is at the core of modern day air traffic control, the standard lacks basic security features such as encryption and authentication. As a result, it is possible to spoof ADS-B data and in the process create unprecedented operational havoc in the skies. In this work we propose FlightSense: a robust adversarial learning based system for filtering out spoofed ADS-B data and subsequent identification of airplanes operating in the airspace from the filtered signal. We use the framework of a generative adversarial network (GAN) for our implementation, which is end-to-end in that it uses the raw I/Q signal data as input and no preprocessing steps are required. We present experiments and results to demonstrate the efficacy of our methods using a real world standardized ADS-B dataset.

Computation of accurate indoor positioning information is important in several areas like mobile robotics, large scale sensor networks, smart city, virtual reality and applications involving internet-of-things (IoT). In spite of its growing importance due to the advent of large scale autonomous deployments of sensor networks and IoTs, we still do not have a solution for indoor positioning that is equivalent to the Global Positioning System (GPS), which is the standard for large scale outdoor 10- calization. However GPS is not useful for indoor positioning as it is often unreliable and/or unavailable in indoor environments. In this paper we present a multi-source radio frequency (RF) based framework for automatic indoor positioning using received signal strength (RSS) and demonstrate its efficacy by simultaneously using broadcast FM radio & GSM signals for position estimation. Our framework is data driven and can be justified using Bayesian minimum risk analysis and is easy to extend by incorporating other sources of RF signals (like WiFi signals) and thus provides a generalized framework for building indoor positioning systems using signals of opportunity. We call our framework RF-MSiP: Radio Frequency Multi-Source Indoor Positioning Using our algorithms with the well known AMBILOC dataset, we can localize exactly for approximately 98.7% of the test locations over a period of one year, which demonstrates not only the efficacy of the algorithms but also its resiliency to change in the RF environment across the year, thus establishing the transfer learning capability of our system.

In this paper we attempt to fuse both airborne Lidar and high resolution images (0.5 feet per pixel) to identify road networks in a large geographic region. We perform pixel-wise segmentation to classify each pixel as road or non-road based on color and depth features in a larger neighborhood context. This constitutes a bimodal setting because the RGB pixels represent the color space and the depth values come from three dimensional Lidar readings. We present multiple strategies for fusing Lidar and images. We describe a cost-effective, modular, deep convolution network design, TriSeg which gives better IoU metric for the aerial road segmentation problem than the state of the art RGB only architectures. We report on many other architectures as well as release our dataset for further research: https://bitbucket.org/biswas/fusion_lidar_images.

Truth Finding is the problem of determining correct information from several conflicting sources and is required for data aggregation. Existing algorithms solve the problem by simultaneously estimating source qualities and fact confidences, working on either numeric or non-numeric data. However, in practice, datasets are a mixture of several different data types. In this work we present a unified framework for finding truth from a collection of conflicting, authoritative sources. We assume that a small subset of independent reliable sources are selected by a preprocessing step. We formulate truth finding as an outlier removal problem, by modeling the similarities between the values reported by these sources. Our algorithm works in two stages: it first models the similarity graph between the sources and then finds the truth by invoking an outlier removal algorithm. We report experiments on several datasets including results for fixing records in Open Library; an open, editable library catalog.

In this paper, we present a coarse and passive localization system for GPS-denied environments. Our system is based on a cheap software defined radio (SDR) costing less than $25, which is used to listen to broadcast signals from local FM Radio stations. We show that the hardware and associated algorithms are capable of localizations with average errors of less than 5 miles, without requiring a fingerprinting or crowd sourcing approach. The algorithm is based on a large-scale simulation based map building phase and a query phase. We use 924 power spectra distributed over more than 200 miles for experimentally validating our query algorithm. Our average localization error is less than 5 miles.