Pratyush Muthukumar

Computer Science M.S. Student
School of Engineering
Stanford University


Generative Adversarial Imitation Learning for Empathy-based AI

Imitation Learning · Adversarial Training · Dialogue Models

  • Utilized the Generative Adversarial Imitation Learning architecture to model the complex, high dimensional state space for natural conversation dialogue models.
  • Implemented text generation through the GPT-2 deep pre-trained language model within the generator of the GAIL model.
  • Evaluated performance against expert empathetic trajectories using perplexity and BLEU score NLP metrics.
  • Paper GitHub


    A Stochastic Time Series Model for Predicting Financial Trends with NLP

    Financial Forecasting · Generative Adversarial Network · Sentiment Analysis

  • Utilized Naive Bayes Sentiment Analysis to generate high-level representations of financial text data such as Earnings Conference Calls (ECCs) and news articles.
  • Utilized textual representations as latent vector for the LSTM generator trained against a CNN discriminator of a generative adversarial network.
  • Our two-stage model effectively forecasts binary trends, rolling month averages, and fixed time horizon close prices of various aerospace stocks and composite indices.
  • Paper

    Real-Time Spatiotemporal Air Pollution Prediction with Deep Convolutional LSTM through Satellite Image Analysis

    Air Pollution Prediction · Spatiotemporal Data Analysis · ConvLSTM

  • Utilized the deep Convolutional Long-Short Term Memory (ConvLSTM) architecture to predict spatiotemporal ground-based NO2 in Los Angeles over time.
  • Utilized spatiotemporal data preprocessing on Sentinel-2 nitrogen dioxide remote-sensing satellite imagery and ground-based monitoring site pollutant data.
  • Paper Website

    Satellite Image Atmospheric Air Pollution Prediction through Meteorological Graph Convolutional Network with Deep Convolutional LSTM

    Spatiotemporal Kriging · Graph Convolutional Network · Atmospheric Modelling

  • Utilized the Graph Convolutional Network architecture to perform spatiotemporal kriging on meteorological weighted directed graphs.
  • Applied unsupervised learning graph representation learning to include high-level meteorological data alongside remote-sensing satellite imagery for spatiotemporal pollution prediction.
  • Paper Website

    Predicting PM2.5 Atmospheric Air Pollution using Deep Learning with Meteorological Data and Ground-based Observations and Remote-Sensing Satellite Big Data

    PM2.5 Prediction · Multisource Big Data · Deep Learning

  • Utilized NASA MAIAC MODIS Aerosol Optical Depth remote-sensing satellite imagery alongside ground-based sensor observations in a deep learning model.
  • Predicted ground-based PM2.5 in micrograms per cubic meter in the greater Los Angeles area and reported unparalled experimental accuracy through RMSE and NRMSE error values.
  • Paper Website