I am pursuing a Master's degree in Computer Science at Purdue University. My research interests lie in applying machine learning and computer vision for efficient image synthesis. I am a member of AAAI, SASE, and SIAM.
Actively looking for full-time opportunities in Deep Learning and Computer Vision

News

  • [Feb ‘21] Paper Submitted to ICML '21 for review
  • [Dec ‘20] Poster Presented at NeurIPS '20 Beyond BackPropagation Workshop
  • [Aug ‘20] Attended Oxford Machine learning Summer School OxML '20
  • [Jul ‘20] Presented and Volunteered at PEARC '20
  • [Jul ‘20] Student Volunteer at ICML '20
  • [Apr ‘20] Paper accepted at PEARC '20

Publications

  • Learning Flows Locally [Under Review]
    M. Bhatt
    , and D. Inouye
  • Learning Flows By Parts, NeurIPS ‘20 Workshop on Beyond BackPropagation: Novel Ideas for Training Neural Architectures [Paper] [Poster] [Video]
    M. Bhatt
    , and D. Inouye
  • Design and Deployment of Photo2Building: A Cloud-based Procedural Modeling Tool as a Service, PEARC ‘20 [Preprint] [Published] [Code] [Web-App]
    M. Bhatt
    , R. Kalyanam, G. Nishida, L. He, C.K. May, D. Niyogi and D.G. Aliaga

Research Experience

Scaling Photo2Building [Paper] [2019]

Procedural Modeling has been used to create 3D models and textures from a set of grammars. The tool Photo2Building procedurally generates a 3D geometry of the building given a photograph and an input silhouette. Our work focused on using cloud infrastructures to make the tool scalable and publicly available for the urban modeling community. Our solution involved developing a client-server architecture where the server used a pre-trained CNN model to estimate the building (facade and window) grammars and the client provided an interactive GUI developed using Jupyter Notebook. Our results showed that it costed only 0.60 USD of Google Cloud resources for reconstructing a single 3D building.

Silhouette Detection [In-Progress] [2020]

The main disadvantage of Photo2Building is that it requires a manually labelled contour as an input. This work focuses on performing an instance segmentation on street-view images to produce accurate 3D polylines representing the building boundary. Our network uses an unsupervised training approach. We compare results by training on well-known instance segmentation models such as Mask R-CNN, YOLACT and Mask Scoring R-CNN.

Learning Generative Normalizing Flows by Parts [Under Review] [2020]

State-of-the-art flow based models introduce significant model complexity in deep neural networks which induces long training times even on high-end GPUs. We focus on optimizing such models in parts rather than full end-to-end backpropagation to see if we can achieve similar performance even without end-to-end backpropagation. Each model part is gradient-isolated to not exchange gradient information with other parts. We evaluate our results along for training the state-of-the-art Glow model via parts where we achieve comparable accuracy in
~30% shorter time
.

Course Projects

Predicting Remaining Useful Life (RUL) of a Li-Ion battery [Report]

The project involved estimating the remaining useful life (or capacity) of a Li-Ion battery after k number of charge and discharge cycles on a dataset provided by NASA. We evaluated various machine learning techniques to generate a user-dependent prediction. Relevance Vector Machine(RVM) with a certain degree of randomness was used to determine the capacity change

Fake News Detection [Report]

Classified news headlines during the U.S. Presidential Elections of 2016 into real and fake. A TF-IDF approach was used to generate lemmatized data. We evaluated our LSTM model against commonly used algorithms available from scikit-learn and achieved an accuracy of 96%

Gerstner Wave Simulation

Developed a 3D water simulation tool based on the Gerstner Wave simulation algorithm on the GPU. The code was developed using OpenGL and FreeGLUT

Work Experience

Senior Software Developer [Oct'17 - Dec'18]

Teaching Assistant [Aug'20 - Present]

CS 180: Problem Solving And Object-Oriented Programming