Metarya Ruparel

Digital images are easy to manipulate and edit due to availability of powerful image processing and editing software. Nowadays, it is possible to add or remove important features from an image without leaving any obvious traces of tampering. As digital cameras and video cameras replace their analog counterparts, the need for authenticating digital images, validating their content, and detecting forgeries will only increase. Detection of malicious manipulation with digital images (digital forgeries) is the main topic of this project. Throughout the years, various computer vision and deep learning approaches have been pro- posed to solve the issue. In particular, a few of the CNN architectures suggested, manage to predict images with an accuracy of more than 96%. That said, the images used in those studies are easily recognized by humans. This raises a crucial question: how would CNNs perform on more challenging samples? In this study, we develop a CNN network inspired by a previous study and answer this question by analyzing various approaches to measure performance on CASIA2 and MICC datasets. We also measure the effect of a data augmentation technique and different hyper-parameters on classification performance.

Python based Machine Learning model which summarizes an event and detects all the sub-events from a corpus of tweets pertinent to that Event. Used various clustering algorithms like LSH, K-means and Hierarchical clustering to detect sub-events. Used a variant of tf-idf, to rank each tweet in a cluster and produced a summary by picking the top tweet from each cluster and sorting them based on their timestamps. I successfully detected the sub-events and produced a summary for FA-Cup Finals 2017.

NYU Grad Project - Research on different Unsupervised AD techniques. This approach also addresses the caveats of unsupervised dataset and how to tackle it. Approaches ranges from statistical to Neural Nets.

Research on different time-series Anomaly Detection Algorithms like RNNs, Hierarchal Temporal Memory, KMeans, etc.

The code contains the implementation of idea of the following paper with our improvements. A statistical based model which, when given a query image finds all the images similar to it from a corpus of images. Used a faster k-means with the k-means++ initialization from scikit’s library. On top of that, created a distributed input division to parallelly execute the k-means algorithm with different configuration across different cores. Used LSH with p-stable distribution to match queries with similar images. Used vector hashing to decrease the turn around time and increase throughput.

Repo for the Machine Common Sense project

A distributed database, complete with multiversion concurrency control, deadlock detection, replication, and failure recovery.

We compare two parallel programming approaches for multi-core systems: the well-known OpenMP and Threading Building Blocks (TBB) library by IntelR . The comparison is made using the paral- lelization of different real-world algorithm like MergeSort, Matrix Multiplication, and Two Array Sum. We develop several parallel implementations, and compare them w.r.t. programmability, scalability, run-time overhead, and the amount of control given to the programmer. We show that TBB requires a considerable program re-design, whereas with OpenMP simple compiler directives are sufficient. While TBB appears to be less appropriate for parallelizing existing implementations, it fosters a good programming style and higher abstraction level for newly developed parallel programs. Our experimental measurements on Two Intel Xeon E5-2680 (2.80 GHz) (20 cores) demonstrate that TBB slightly outperforms OpenMP in our implementation.

No description.

Notes for 2020.

Repository for final project proposals

A Python implementation of global optimization with gaussian processes.

The Natural Language Decathlon: A Multitask Challenge for NLP

Production-Grade Container Scheduling and Management

TransmogrifAI (pronounced trăns-mŏgˈrə-fī) is an AutoML library for building modular, reusable, strongly typed machine learning workflows on Spark with minimal hand tuning