Biswarup Bhattacharya

An important means of controlling recurrent infectious diseases is through active screening to detect and treat patients. Disease detection on a large network of individuals is a challenging problem, as the health states of individuals are uncertain and the scale of the problem renders traditional dynamic optimization models impractical. We propose a novel active screening model (ACTS) and an algorithm (TRACE) to facilitate active screening for recurrent diseases: a new belief-state approach to modeling SEIS type diseases, and a community and eigenvalue-based algorithm to generate an online policy for multi-round active screening. We demonstrate applicability through extensive experiments on several real-world datasets emulating human contact. To the best of our knowledge, this is the first work on a multi-round active screening model and algorithm for diseases with a latent stage and no permanent immunity.

Distributing public health services is a major challenge: limited health workers are responsible for large populations with varied health problems. We propose a hierarchical visitation policy design, scalable to regions of various sizes and diversities, with two levels of planning: macro-level (region-level) by adapting the p-functional regions problem, and micro-level (village-level) via a restless multi-armed bandit model using POMDPs with the Whittle Index Policy. We also address heterogeneity of health problems across villages and the dynamic nature of public health priorities. Preliminary experiments show promising results for health policy planning.

Global average pooling (GAP) has been used previously to generate class activation maps. AdGAP builds on the fact that convolutional filters possess position information of essential features, so combining feature maps can locate class instances in an image. The architecture generates promising results and, unlike previous methods, is not sensitive to input image size, promising wider application.

Autoencoders are essential in learning representations of large data for dimensionality reduction. Images are converted to sparse domain using transforms like FFT or DCT where information requiring encoding is minimal. By optimally selecting feature-rich frequencies, latent vectors are learned more robustly. We show enhanced autoencoder performance in sparse domain for images.

Assessing infection spread accurately is important for effective mitigation and treatment planning. We visualized how contact networks react to introduced infections and identified correlations between predicted health states and actual data using social media. We developed an SEIRS disease model for realistic simulation, optimized parameters using historical data from Twitter and Flickr, predicted individual health states, and evaluated on a Los Angeles network for the 2018 flu season.

Using the Global Appraisal of Individual Needs (GAIN) dataset covering adolescents and emerging adults recovering from substance abuse, we (1) predict whether an individual will relapse given the treatment received and demographic features, and (2) analyze treatment effectiveness across demographic variables. The eventual goal is prescribing personalized treatments to maximize likelihood of recovery and minimize relapse.

Recommender systems often provide recommendations without understanding the actual intent of the user. We built a system that tracks user activity on a web application and determines session intent: user activity is encoded per session, information seen by the user is represented in high dimensions and reduced via tensor factorization (where intent scoring occurs), and activity plus context yields a recommendation score, ranked with filtering and collaborative techniques. Experiments on real-world timestamped user activity data — recommending reports in a business analytics tool — beat the baselines.

Non-availability of reliable and sustainable electric power is a major problem in the developing world, and solar faces uncertainties in weather, storage, and land use, plus issues like power mis-commitment and congestion. We propose a deep learning-based system for predicting faults and selecting power generators optimally to reduce costs and improve reliability in solar power systems, with highly encouraging results.

A malfunction of even a small part of a power grid can cause loss of productivity, revenue, and in some cases life. We built a system that analyzes grid information at any given time and determines grid health using formal models and machine learning techniques like recurrent neural networks. Grid conditions — faults, generator output fluctuations, load fluctuations — are simulated with Siemens PSS/E, trained with classifiers like SVM and LSTM, and tested with very high accuracy. The model scales to larger and more complex grid architectures.

Building on Global Average Pooling for class activation, SIMILARnet exploits the position information in convolutional filters to locate class instances in an image. We propose a biologically inspired model free of differential connections that requires no separate training, reducing computation overhead. The model is not sensitive to input image size, promising wider application.

Optimal generator subset selection requires evaluating source type, generation, transmission and storage capacities, and congestion. We simulated grid conditions including generator supply, weather, and load demand with Siemens PSS/E, trained deep learning models on this data, and tested with highly encouraging results — to our knowledge the first working, scalable deep learning model for this problem.

We introduce a model that scores an ATM location as an indicator of its relative likelihood of transactions. Two concurrent prediction models capture spatially dynamic features: a local model encoding spatial variance from highly energetic features, and a global model enforcing dominant trends across the data as feedback to prevent overfitting. Lacking an objective function, the model is trained on a synthetic objective built from dominant k-means features in the local model, with encouraging results.

An intelligent system learns the handwriting of an entity using GANs, via a modified DCGAN architecture, with reinforcement learning techniques discussed for faster learning. Applications include identification of forged documents, signature verification, computer-generated art, and digitization of documents. Early implementation shows good performance on MNIST.

We approach learning synthetic driving using generative neural networks: a controller trainer network using images plus key-press data to mimic human learning. A stable GAN architecture predicts between driving scenes using key presses. The model is trained on Road Rash and tested on other maps to determine the extent of learning.