Shreesh Ladha

I am a Machine Learning Engineer at Facebook working on building NLP based automated solutions for advertisers. Previously, I was a graduate student in the department of Computer Science at University of Massachusetts Amherst.

Before coming here, I worked for a year at Samsung Research Institute, Bangalore in their Bixby NLU Research division. My work revolved around intent recognition and entity extraction. Prior to that, I did my undergraduation from Indian Institute of Technology Kanpur with a major in Mathematics and Scientific Computing.

Email  /  LinkedIn  /  Resume

Single Image Super Resolution using CNN's
Supervisor: Dr. Subhransu Maji, University of Massachusetts Amherst

Implemented a fully convolutional net based on the ResNet architecture for transforming a Low Resolution (LR) image to High Resolution (HR).

Experimented with learning methods to additionally improve the resolution by providing an additional HR image similar to the LR image during training

YouTube-8M Video Understanding Challenge

As part of a Kaggle competition, experimented with different CNN architectures (in Tensorflow) for adaptive pooling of frames within a Youtube video for classification task (with constraints on model size).

Finished in the top 12% in the competition, comprising of close to 400 teams. All model training was performed on Google Cloud's ML Engine.

Cross Lingual Embeddings for POS tagging
Supervisor: Dr. Brendan O'Connor, University of Massachusetts Amherst

Explored the possibility of transferring information from high resource languages such as English to improve the performance of POS-Taggers for languages with low resources, in our case, Hindi in a completely semi-supervised way.

Used the tags obtained in this fashion in multiple auxiliary tasks and obtained significant improvement in accuracies.

SeeDB : Efficient Data-Driven Visualization

Implemented SeeDB : a visualization recommendation engine to facilitate fast visual analysis paper from scratch.

SeeDB intelligently explores the space of visualizations, evaluates promising visualizations for trends, and recommends those it deems most useful or interesting (by observing large deviations from some reference).

A survey of Zero Shot Learning
Supervisor: Dr. Piyush Rai, Indian Institute of Technology Kanpur

Studied different methods of performing Zero Shot Learning(ZSL) - prediction of a label that has been not seen during the training procedure.

Implemented two contemporary papers from this area which required learning a common semantic space for embedding images and labels, to perform ZSL task. Focused on dictionary learning as a way to resolve the PDS issue and found that CNN based features drastically improve the classification accuracy.

Visualization of high dimensional data
Supervisor: Dr. Ketan Rajawat, Indian Institute of Technology Kanpur

Explored applications of convex optimization for dimensionality reduction, especially over non linear manifolds.

Compared performance based on visualizations, computational complexities, and error rates obtained in classification tasks. Selected as the best project in the course comprising of over 80 students.

OCR of conjunct characters in Devanagari Script
Supervisors: Dr. Harish Karnick, Indian Institute of Technology Kanpur
Dr. Amit Mitra, Indian Institute of Technology Kanpur

Devised algorithms for conjunct character recognition within scanned documents. Experimented with pre-trained and self-trained CNN architectures.

Obtained significant improvement over preexisting systems with a decrease in average word error rate from 19.5% to 12.5%.

Domain-invariant Transfer Kernel Learnning
Supervisor: Dr. Harish Karnick, Indian Institute of Technology Kanpur

Implemented a learning model which generalizes across training and testing data from different distributions.

Minimizing the Nystrom Approximation error, obtained a domain-invariant kernel which is plugged into an SVM for transfer learning.

Anomaly detection in Time Series
Supervisor: Dr. Amit Mitra, Indian Institute of Technology Kanpur

Developed a framework to detect amplitude anomalies and shape anomalies within a temporal data over a time series. An autocorrelation representation of the time series was employed to capture the shape information

Samsung Research and Development Institute, Bengaluru (SRIB)
Supervisor: Singaravel Ramalingam, Principal Engineer, SRIB

Created a system for analyzing and suggesting improvements in Samsung's voice assistant.

Used Apache Spark to work with terabytes of data and applied classification and clustering algorithms for better insights. Wrote queries to gather insights using SQL and Map-Reduce code, in Spark.

m.Paani, Mumbai
Supervisor: Akanksha Hazari, Founder & CEO

m.Paani is a Hult-Prize winning social startup working on loyalty programmes for people living at the bottom of the pyramid. Built an interactive map, for spatial data analytics, using web technologies and open source javascript libraries.

Presented the above tool to a CEO of a large company that is an m.Paani partner. Designed the front-end interfaces of m.Paani's loyalty management application.

inspired from this website