Shubham Dash
Grad Strudent at University of Michigan, Ann Arbor
Ann Arbor, Michigan
shubhamd@umich.edu
(510)-378-2995
Skills
Data Structures and Algorithms
80%
Research
70%
Deep Learning and Machine Learning
75%
Languages
English
Hindi
German
Experience
Research Scientist Intern@Schlumberger Doll Research, Cambridge, MA
May 2018 - Sep 2018
- Working on building deep learning models to estimate the materials surrounding in an oil well, and analyze the well integrity using this information.
- Using acoustic data in the form of sonic and ultrasonic data as well as spectrogram like images to build the models
- Using state of the art techniques adapted from ICML 2017, 2018 and ICLR 2017 in the form of GANs and Variational Autoencoder for building generative models
- Working towards filing a patent encompassing the novel deep learning used for the application.
Technologies used
- Python 2.7+
- Tensorflow 1.7.0+
- Keras 2.2.0+
- MATLAB / Octave (for visualization)
Machine Learning Intern@Indian Institute of Technology, Madras, Chennai, India
June 2016 - Aug 2016
- Developed a model based on Natural Language Processing, Machine Learning and Deep Learning to gauge the strength of the sentiment using textual features.
- The techniques involved in building the model were ordinal regression and Deep Learning methods such as Recurrent Neural Networks
- The dataset was collected-compiled-cleaned by me and contained 2000+ reviews scraped from Google Play Store
Technologies used
- Python 2.7+
- Selenium Webdriver
- R
- Keras 2.2.0+
- Beautiful Soup
Software Intern@Snapdeal.com
Jun 2010 - Mar 2012
- Developed a scraper to automate the process of categorizing the advertisements apperaing on Goolge Search results.
- The advertisements were categorized into classes like organic search results, Product Listed Advertisements and SEMs and helped improve the brands visibilty on Google.
- Created a communication channel between the company and the customers using third-party Whatsapp APIs and SQL database that stores consumer queries.
Technologies used
- Python 2.7+
- Selenium Webdriver
- SQL
- Yowsup
- Beautiful Soup
Education
University of Michigan, Ann Arbor: GPA : 3.9 / 4.0
Sep 2017 - Apr 2019
Masters of Sciences, specializing in Signal and Image Processing and Machine Learning
Courses:
- Probability and Random Processes
- Linear Algebra for Machine Learning
- Machine Learning
- Theory of Machine Learning
- Computer Vision
- Large Graph Mining
Netaji Subhas Institute of Technology
Aug 2013 - June 2017
Bachelors of Engineering, Electrical Engineering, came 1st in a class of 55 students with a department rank of 8/120, scored 81.02% overall.
Courses:
- Data Structures and Algorithms
- Artificial Intelligence
- Digital Signal Processing
- Control Systems
- Numerical Methods for Computing
Projects
Jan 2018 - Apr 2018
Project carried out under the supervision of Prof. Mert Pilanci
- Developed methods under the idea of sketching to optimize the performance of deep networks
- Tested the theory by employing it on Convolutional Neural Networks over MNIST and CIFAR-10 datasets
- The performance of a "pruned" network was almost similar to the orginal model and at the same time sppeding up the process multi fold.
Hybrid GAN structure with Mutual Info GAN, Wasserstein GAN and Gradient Penalty
Ongoing
- The Hybrid GAN structure is used for unsupervised learning and clustering based on hybrid of state-of-art GAN structures inspired from latest ICLR and ICML publications
- Tested the theory by employing it on Convolutional Neural Networks over MNIST and CIFAR-10 datasets
Contextual Analysis of Videos based on Graph Representaion
Jan 2018 - Apr 2018
- Developed a graph based model to answer contextual questions based on videos
- The model was built on Stanford CoreNLP engine, and object detection dataset and pretrained model.
Pose Estimation for Activity Detection
Mar 2018 - Apr 2018
- The project involved using CNN based posenet to estimate the pose of a pedestrian in order to detect their activity.
- Support Vector Machine and Fuzzy Logic used to classify the actions into classes such as Pedestrian looking down, talking on phone and looking away.
- Built the dataset ourselves by recording videos for specific actions and processing on a frame level.
Multi Label Classification for board game classification
Sep 2017 - Dec 2017
- Built classification models using statistical methods and CNNs, which could classify a board game into different genres using physical and image features to build the model.
- Statistical models involved using Binary Relevance and Classifier Chains
- The dataset was obtained from Kaggle, containing more than 5000 games, with 84 features involving mechanical and physical features of games as well meta data
- Also implemented CNN based classifier based on the cover art which is also available to us with the dataset.