Created a system that can generate captions from an image and can recommend similar images from a dataset or Google images. The system uses an image captioning network based on Resnet 152 encoder and a single layer LSTM based decoder. This system can be used to create an entire dataset of images, from few seed images.
Used various machine learning algorithms like Convolutional Neural Networks, Multi Layer Perceptrons, AbaBoost and Random Forest to implement a system to evaluate handwritten mathematical expressions. The system works by locating and identifying digits and math symbols from an image of handwritten mathematical expressions, and processing them to evaluate the expression.
Designed a system to analyze and visualize the IMDB dataset that contain information about 4.3 million titles and 8 million artists. The system can be used to obtain hard to collect information like popularity of a particular genre or a director throughout the year, most popular movies of a particular year, average rating and popularity of a particular genre etc.Also designed a movie recommender system based on collaborative filtering.
Designed a classification model that checks whether is news headline is fake or not. The model uses n-grams, tf-idf, non-negative matrix factorization and Latet Dirichlet Allocation to check if a news headline is similar to that of fake news.
Designed a recommender system for suggesting businesses to customers. For this task, used google local business data which had over 200,000 reviews from 20,000 users, for 19,000 businesses. Used various techniques like matrix factorization, collaborative filtering and latent factor models to predict if a user would visit a business and the likely rating the user would give to that business.
Designed a word suggestion program using n-gram mixture models. The models tries to suggest next likely word, given a sequence of words using a weighted combination of a unigram, a bigram and trigram model and was trained using text data from various open sources.
Implemented a portfolio selection algorithm, using Cover's universal portfolio theory. The algorithm works by trying to track the best possible constant rebalance portfolio, on a set of stocks. To run the algorithm, stocks from 14 companies were downloaded from Yahoo! finance.
Developed an agent to play rock paper scissors, using the concept of universal probability. The agent tries to estimate the underlying probability distribution of the opponent to defeat the most probable move.