Mobile radiation detectors aim to help identify sources of radiation. Finding a radioactive sourc... more Mobile radiation detectors aim to help identify sources of radiation. Finding a radioactive source in a man-made environment such as a city can be challenging because the additive signal received by the detector contains both photon counts from the source of interest and from the cluttered and variable ambient background. Decomposing the overall radiation spectrum into its background and source signal components is key. When either or both of the background or source components is low in photon counts, the estimation of signal components can become especially challenging.
Gamma-ray spectrometry data is typically presumed to be created with a Poisson process, though Gaussian-based estimators are typically used to approximate the truly Poisson-distributed data. Generally this approximation suffices, but performance loss can occur when photon counts affecting a sensor are low in number for any signal component. Low photon count signal and/or noise components can occur in a variety of real world scenarios. Photon counts from the source may be low because the source is very weak or only observable from large standoff distances. Photon count rate from both source and background may be low if small sensors (with limited surface area) are used or if measurement time is limited.
Our study experiments with augmenting established anomaly detection and match filter signal component estimators with Poisson-based models. We apply estimators such as the Poisson Principal Component Analysis (Poisson PCA) and the Zero-Inflated Poisson (ZIP) models to the source detection problem and benchmark with respect to popular estimators in the literature. Finally, we apply Bayesian Aggregation to the Poisson-based estimators to aggregate evidence across multiple spatially-correlated sensor observations. Our results indicate that the use of such techniques can aid threat detection when photon counts are low in signal and/or background noise components.
Mobile radiation detector systems aim to help law enforcement officers effectively identify the k... more Mobile radiation detector systems aim to help law enforcement officers effectively identify the key sources of radiation in environments (such as urban scenes) while minimizing frequency of false alarms caused by non-threatening nuisance sources. In this study, we develop methods for spatially aggregating evidence from multiple spectral observations to locate and infer properties of threatening radiation sources.
Our Bayesian Aggregation (BA) framework uses simulation and non-parametric statistical techniques to learn distributions for expected Signal-to-Noise (SNR) scores of empirical background radiation measurements, either with or without simulated injection of threatening point sources. Information from multiple measurements is combined via Bayesian update rules which aggregate evidence by modeling multiple measurements as conditionally independent given the underlying hypothesis about the source. Using an optimized algebraic formulation and efficient data structures, BA is much more scalable than naive posterior distribution calculation.
Our framework supports modeling of parameters such as source intensity and source type to enable detection and characterization of faint sources of various types. We compare different methods of incorporating this information in the BA hypothesis space and benchmark our methods against existing techniques of aggregating multiple observations to detect sources.
Lunar Tabs is an intelligent screen reader friendly application that makes the millions of guitar... more Lunar Tabs is an intelligent screen reader friendly application that makes the millions of guitar tabs online accessible to persons with low vision. An accessible user interface is designed for different use scenarios including for desktop platforms and mobile devices. Methods are developed to convey repetition information in music effectively over a screen-reader, and capabilities to use the application in several hands-free modes are prototyped. User studies are conducted to evaluate the technology’s potential impact in the blind/low vision community.
The copious amounts of freely available guitar tablature (tab) online pose interesting Machine Le... more The copious amounts of freely available guitar tablature (tab) online pose interesting Machine Learning challenges. Difficulty Level Classification, the problem of automatically rating the difficulty of a guitar tab, is an important challenge in helping users find appropriate musical pieces to autonomously improve their guitar skills. In contrast to previous approaches, our rating scheme allows the user to provide labeled data in order to forecast ratings that are personalized for the user, rather than relying on a universal ranking of difficulty. We show how feature selection can be used in conjunction with classification to build robust models of tab difficulty for individual users. The top features used in prediction are also presented to the user to help them diagnose their needs for improvement. Finally, we prototype Active Learning techniques to reduce the labeled sample size required from the user.
The human brain’s control strategies for using a tool such as a hammer show promise for improving... more The human brain’s control strategies for using a tool such as a hammer show promise for improving tool use on robotic arms. The specific case of teaching a robotic arm to drive a nail into a board is utilized in this study to evaluate two biologically-inspired controller designs for the ballistic swing of a hammer. The study compares a Q-Learning Table Controller and a Population Coded Neural Network Controller and presents simulation results. In both cases, the proposed controller designs learn to improve their efficiency at driving nails into the board as they drive more nails.
Mobile radiation detectors aim to help identify sources of radiation. Finding a radioactive sourc... more Mobile radiation detectors aim to help identify sources of radiation. Finding a radioactive source in a man-made environment such as a city can be challenging because the additive signal received by the detector contains both photon counts from the source of interest and from the cluttered and variable ambient background. Decomposing the overall radiation spectrum into its background and source signal components is key. When either or both of the background or source components is low in photon counts, the estimation of signal components can become especially challenging.
Gamma-ray spectrometry data is typically presumed to be created with a Poisson process, though Gaussian-based estimators are typically used to approximate the truly Poisson-distributed data. Generally this approximation suffices, but performance loss can occur when photon counts affecting a sensor are low in number for any signal component. Low photon count signal and/or noise components can occur in a variety of real world scenarios. Photon counts from the source may be low because the source is very weak or only observable from large standoff distances. Photon count rate from both source and background may be low if small sensors (with limited surface area) are used or if measurement time is limited.
Our study experiments with augmenting established anomaly detection and match filter signal component estimators with Poisson-based models. We apply estimators such as the Poisson Principal Component Analysis (Poisson PCA) and the Zero-Inflated Poisson (ZIP) models to the source detection problem and benchmark with respect to popular estimators in the literature. Finally, we apply Bayesian Aggregation to the Poisson-based estimators to aggregate evidence across multiple spatially-correlated sensor observations. Our results indicate that the use of such techniques can aid threat detection when photon counts are low in signal and/or background noise components.
Mobile radiation detector systems aim to help law enforcement officers effectively identify the k... more Mobile radiation detector systems aim to help law enforcement officers effectively identify the key sources of radiation in environments (such as urban scenes) while minimizing frequency of false alarms caused by non-threatening nuisance sources. In this study, we develop methods for spatially aggregating evidence from multiple spectral observations to locate and infer properties of threatening radiation sources.
Our Bayesian Aggregation (BA) framework uses simulation and non-parametric statistical techniques to learn distributions for expected Signal-to-Noise (SNR) scores of empirical background radiation measurements, either with or without simulated injection of threatening point sources. Information from multiple measurements is combined via Bayesian update rules which aggregate evidence by modeling multiple measurements as conditionally independent given the underlying hypothesis about the source. Using an optimized algebraic formulation and efficient data structures, BA is much more scalable than naive posterior distribution calculation.
Our framework supports modeling of parameters such as source intensity and source type to enable detection and characterization of faint sources of various types. We compare different methods of incorporating this information in the BA hypothesis space and benchmark our methods against existing techniques of aggregating multiple observations to detect sources.
Lunar Tabs is an intelligent screen reader friendly application that makes the millions of guitar... more Lunar Tabs is an intelligent screen reader friendly application that makes the millions of guitar tabs online accessible to persons with low vision. An accessible user interface is designed for different use scenarios including for desktop platforms and mobile devices. Methods are developed to convey repetition information in music effectively over a screen-reader, and capabilities to use the application in several hands-free modes are prototyped. User studies are conducted to evaluate the technology’s potential impact in the blind/low vision community.
The copious amounts of freely available guitar tablature (tab) online pose interesting Machine Le... more The copious amounts of freely available guitar tablature (tab) online pose interesting Machine Learning challenges. Difficulty Level Classification, the problem of automatically rating the difficulty of a guitar tab, is an important challenge in helping users find appropriate musical pieces to autonomously improve their guitar skills. In contrast to previous approaches, our rating scheme allows the user to provide labeled data in order to forecast ratings that are personalized for the user, rather than relying on a universal ranking of difficulty. We show how feature selection can be used in conjunction with classification to build robust models of tab difficulty for individual users. The top features used in prediction are also presented to the user to help them diagnose their needs for improvement. Finally, we prototype Active Learning techniques to reduce the labeled sample size required from the user.
The human brain’s control strategies for using a tool such as a hammer show promise for improving... more The human brain’s control strategies for using a tool such as a hammer show promise for improving tool use on robotic arms. The specific case of teaching a robotic arm to drive a nail into a board is utilized in this study to evaluate two biologically-inspired controller designs for the ballistic swing of a hammer. The study compares a Q-Learning Table Controller and a Population Coded Neural Network Controller and presents simulation results. In both cases, the proposed controller designs learn to improve their efficiency at driving nails into the board as they drive more nails.
Uploads
Papers by Prateek Tandon
Gamma-ray spectrometry data is typically presumed to be created with a Poisson process, though Gaussian-based estimators are typically used to approximate the truly Poisson-distributed data. Generally this approximation suffices, but performance loss can occur when photon counts affecting a sensor are low in number for any signal component. Low photon count signal and/or noise components can occur in a variety of real world scenarios. Photon counts from the source may be low because the source is very weak or only observable from large standoff distances. Photon count rate from both source and background may be low if small sensors (with limited surface area) are used or if measurement time is limited.
Our study experiments with augmenting established anomaly detection and match filter signal component estimators with Poisson-based models. We apply estimators such as the Poisson Principal Component Analysis (Poisson PCA) and the Zero-Inflated Poisson (ZIP) models to the source detection problem and benchmark with respect to popular estimators in the literature. Finally, we apply Bayesian Aggregation to the Poisson-based estimators to aggregate evidence across multiple spatially-correlated sensor observations. Our results indicate that the use of such techniques can aid threat detection when photon counts are low in signal and/or background noise components.
Our Bayesian Aggregation (BA) framework uses simulation and non-parametric statistical techniques to learn distributions for expected Signal-to-Noise (SNR) scores of empirical background radiation measurements, either with or without simulated injection of threatening point sources. Information from multiple measurements is combined via Bayesian update rules which aggregate evidence by modeling multiple measurements as conditionally independent given the underlying hypothesis about the source. Using an optimized algebraic formulation and efficient data structures, BA is much more scalable than naive posterior distribution calculation.
Our framework supports modeling of parameters such as source intensity and source type to enable detection and characterization of faint sources of various types. We compare different methods of incorporating this information in the BA hypothesis space and benchmark our methods against existing techniques of aggregating multiple observations to detect sources.
Gamma-ray spectrometry data is typically presumed to be created with a Poisson process, though Gaussian-based estimators are typically used to approximate the truly Poisson-distributed data. Generally this approximation suffices, but performance loss can occur when photon counts affecting a sensor are low in number for any signal component. Low photon count signal and/or noise components can occur in a variety of real world scenarios. Photon counts from the source may be low because the source is very weak or only observable from large standoff distances. Photon count rate from both source and background may be low if small sensors (with limited surface area) are used or if measurement time is limited.
Our study experiments with augmenting established anomaly detection and match filter signal component estimators with Poisson-based models. We apply estimators such as the Poisson Principal Component Analysis (Poisson PCA) and the Zero-Inflated Poisson (ZIP) models to the source detection problem and benchmark with respect to popular estimators in the literature. Finally, we apply Bayesian Aggregation to the Poisson-based estimators to aggregate evidence across multiple spatially-correlated sensor observations. Our results indicate that the use of such techniques can aid threat detection when photon counts are low in signal and/or background noise components.
Our Bayesian Aggregation (BA) framework uses simulation and non-parametric statistical techniques to learn distributions for expected Signal-to-Noise (SNR) scores of empirical background radiation measurements, either with or without simulated injection of threatening point sources. Information from multiple measurements is combined via Bayesian update rules which aggregate evidence by modeling multiple measurements as conditionally independent given the underlying hypothesis about the source. Using an optimized algebraic formulation and efficient data structures, BA is much more scalable than naive posterior distribution calculation.
Our framework supports modeling of parameters such as source intensity and source type to enable detection and characterization of faint sources of various types. We compare different methods of incorporating this information in the BA hypothesis space and benchmark our methods against existing techniques of aggregating multiple observations to detect sources.