Publications

Bayesian structure learning with generative flow networks

Published in Uncertainty in Artificial Intelligence (UAI), 2022

In this work, we propose to use a Generative Flow Networks (Gflownet, a novel class of probabilistic models) as an alternative to MCMC for approximating the posterior distribution over the structure of Bayesian networks, given a dataset of observations.

Download here

A priori guarantees of finite-time convergence for Deep Neural Networks

Published in arXiv preprint arXiv:2009.0750, 2020

In this paper, we perform Lyapunov based analysis of the loss function to derive an a priori upper bound on the settling time of deep neural networks. Drawing from the advances in analysis of finite-time control of non-linear systems, we provide a priori guarantees of finite-time convergence in a deterministic control theoretic setting by formulating the supervised learning framework as a control problem.

Download here

Narrow-band signal classification using Deep Convolutional Neural Networks

Published in 2018 2nd URSI Atlantic Radio Science Meeting (AT-RASC), 2018

With vast amounts of data collected every night by telescopes like Allen Telescope Array, there is a need for a real time system that can accurately distinguish between signals of interest from radio frequency interference. In this paper, we present our approach based on Convolutional neural networks for classification of narrow band signals. By converting the radio signals into 2D spectrogram image, the problem of signal classification can be transformed to as image classification problem. We demonstrate the effectiveness of CNN technique by using a simple Alexnet network with 5 convolutional layers. We also extend our work to detect signals with multiple labels simultaneously.

Deep Convolutional Neural Networks for the Generation of High Fidelity Images from Radio Interferometer Visibility Data

Published in 2018 2nd URSI Atlantic Radio Science Meeting (AT-RASC), 2018

In this paper, we look at the problem of construction of high fidelity images from stored correlator visibility files in radio astronomy. This work is motivated by the amazing recent successes of CNN autoencoders developed for photographic image processing in the machine learning community. The second half of a photographic autoencoder has a structure very similar to what would be required for image generation from raw radio interferometer data (visibilities). In this paper, we argue that a deep convolutional neural network (CNN) can be a highly (computationally) effective approach to radio interferometer image generation starting from raw visibilities and producing high fidelity, cleaned (deconvolved) images as an output. We consider the linear and nonlinear operations performed in image generation and how they have analogs in a standard CNN. We also discuss the potential computational cost savings that might be had by replacing our complicated image processing pipelines with neural networks.

Download here

Detecting Fast Radio Bursts Using Convolutional Neural Networks

Published in 2018 2nd URSI Atlantic Radio Science Meeting (AT-RASC), 2015

Fast Radio Burst (FRB) are high intensity bursts of radio energy lasting only a few milliseconds with large dispersion measures, suggesting extragalactic origin [1]. FRBs have gained intense interest in the astronomical community because their physical origin is not well understood, and because they may be useful as a probe of the intergalactic medium. We use deep neural networks to distinguish between Fast Radio Bursts and other astronomical or interfering signals.

Download here

Heart rate estimation from non-cardiovascular signals using slope sum function and Teager energy

Published in 2015 International Conference on Industrial Instrumentation and Control (ICIC), 2015

Cardiovascular physiological signals such as ECG and ABP provide direct measure of heart rate. But ECG & ABP recorded in the intensive care unit (ICU) are often seriously corrupted by noise and missing data, which lead to errors in Heart Rate (HR) estimation and incidences of false alarm from ICU monitors. Cardiac activity, because of its relatively high electrical energy, introduces ECG artifacts in non-cardiovascular physiological signals namely EEG, EOG and EMG recordings. This paper presents a HR estimation method from non-cardiovascular signals by detection of R-peaks of ECG artifacts in these signals using slope sum function, energy based function, and a novel Signal Quality Index (SQI) assessment technique. SQIs of non-cardiovascular physiological signals are obtained by correlation of Teager-Kaiser energy (TKE) of these signals with TKE of either ECG or ABP signal. The proposed method gives an accurate estimation of HR from non-cardiovascular signals and is quite useful when ECG or ABP signal are either corrupt or missing.

Download here

ECG artifacts detection in noncardiovascular signals using Slope Sum Function and Teager Kaiser Energy

Published in 2015 International Conference on BioSignal Analysis, Processing and Systems (ICBAPS), 2015

In this work, a new method for ECG artifacts detection from noncardiovascular physiological signals namely EEG, EOG and EMG, without the need of any additional synchronous ECG channel, is proposed. This ECG artifacts (R peaks) detection method uses Slope Sum Function and Teager Kaiser Energy operator with an adaptive threshold. The algorithm has shown improved ECG artifacts detection results as compared to that of direct application of Teager Kaiser energy operator on noncardiovascular signals.

Download here

Voice Activity Detection using Temporal Characteristics of Autocorrelation Lag and Maximum Spectral Amplitude in Sub-bands

Published in Proceedings of the 11th International Conference on Natural Language Processing, 2014

A robust voice activity detection (VAD) is a prerequisite for many speech based applications like speech recognition. We investigated two VAD techniques that use time domain and frequency domain characteristics of speech signal. The temporal characteristic of the autocorrelation lag is able to discriminate speech and nonspeech regions. In the frequency domain, peak value of the magnitude spectrum in different sub-bands is used for VAD.

Download here