Academic-projects

We provided aid to the deep learning programmers to quantify uncertainty in the model hyperparameters and help them to make informed decision while initializing hyperparameters. We have leveraged a first order type Uncertain to represent the uncertainty in the random hyperparameters and choose the best value by performing statistical tests on the distribution.

I have extended Lambda calculus using Coq proof assistant to implement a core language that includes two Python language features (compound comparison statement and generator function). I have also described the syntax, operational semantics and type system of the features.

Classified text documents from a large pool using machine learning techniques. A Naive Bayes classifier was built to classify approximately 20,000 newsgroup documents.

Given millions of documents, for each file, rank the similar documents. I have preprocessed the documents, created weighted word vectors and then implemented Simhash (locality sensitive hashing algorithm to evaluate approximate cosine similarity) to generate 64-bit fingerprint of each document. Finally block permuted hamming search was implemented in the fingerprint space to find the near duplicate.

PEKS with Keyword Search is one of the most used method to search keywords over encrypted data. However, semantic security is not preserved and dictionary attack can help attackers to guess keywords. This project solves that issue implementing PEKS and searching keywords over Bloom Filter.

While using cloud platform or any other insecure memory, attack can be made using the access pattern. Oblivious RAM is the way to hide the memory access pattern with some extra bandwidth and memory overhead. Path ORAM uses a binary tree to store all memory blocks and changes the location of the blocks repeatedly to access the whole branch for a single block.

From a given snapshot of a social network database, we can predict whether a person can be potentially connected to another person, by analyzing existing links. We take two datasets (Facebook dataset from Stanford Large Network Dataset Collection and bibliography dataset from DBLP) and import that into MySQL, and Neo4J (Graph based DB) to evaluate the metrics for different network topology.

This is a complete solution for private practitioner doctors. The doctors can manage patients appointment automatically, generate electronic prescriptions and store patients database. The patient side app helps to schedule appointment, get notification and view medical history.

“Tank Battle” is a game where you have to shoot at the enemy tanks at a high speed in different angle to destroy and continue. Enemies attack in a higher speed and try to enter passing the gamer’s tank.