Rajat Agarwal

Spring '17

Implemented zero shot classification in TensorFlow for two classes of CIFAR-10 dataset by training only on the remaining 8 classes. This was done using shared representation for words and images by mapping images to their GloVe word vectors using a CNN as in Socher et al. (2013) and Zero-Shot Learning by Convex Combination of Semantic Embeddings .

Spring '17

Generated adversarial examples in TensorFlow for a MNIST classifier using fast gradient sign method as in Explaining and Harnessing Adversarial Examples (2015).

Library for ANN

Spring '17

Implemented a library in Python with backpropagation for Fully Connected Layers supporting multiple activation and loss functions for ANN using computational gate approach.

Semantic Segmentation for RBCs

Spring '17

Implemented Learning Deconvolution Network for Semantic Segmentation in Keras to generate a binary mask to identify pixels forming Red Blood Cells in a given input image.

Jan '17

Implemented Deep Convolutional Generative Adversarial Networks (DCGAN) in Keras to generate handwritten digits by training the discriminator on MNIST.

Dec '16

Implemented A Neural Algorithm of Artistic Style (2015) for style transfer from a style image to a target content image in TensorFlow using VGG-16 weights.

Fall '16

Trained a CNN in TensorFlow for localization and detection of 24 alphabets in American Sign Language in a camera input. Accuracy of 99% on localization (IoU >= 0.5) and 98% on top-5 classification on test data - the highest among 15 teams in machine learning class of Fall 2016.

Fall '16

Trained a Deep Q-learning Network Reinforcement Learning agent in TensorFlow to solve CartPole problem in the OpenAI Gym environment.

Vision Classification

Fall '16

Implemented standard papers in TensorFlow to achieve bechmarks results on the CIFAR-10, Google SVHN and MNIST dataset.

June '15

Developed a bot to play the game of Connect4 against a user. Implemented in Java using Minimax tree and state evalutaion score.

Spring '16

I started AutolabJS as an open source project out of college. It is a real time auto evaluation software for programming labs. The software currently supports automatic evaluation in Java, Python2, Python3, C++ and C programming languages. Autolab uses Gitlab as a component to provide the version control system. All the student code submissions reside in Gitlab, thus benefiting from powerful git system. It evaluates the submissions in a distributed environment with load balancing and the software is modular with separate microservice instances for all the application components. The initial release was deployed for OOP course (Java) with 170+ students in Fall 2016 and for multiple courses in future semesters.
Check the project on GitHub if you wish to use deploy it for your labs or contribute.

Spring '17

Implemented using YACC, it uses LALR(1) bottom up parsing to seperate fields and values in the given input BibTeX code.

Spring '17

It uses Flex to convert input C code to HTML and shows it the browser with line numbers and syntax highlighting.

FTP Client and Server

Spring '17

Implemented in C using socket programming, it supports FTP operations including directory browsing.

Multiuser chat service

Spring '17

Implemented a chat service in C, where users can send unicast, multicast and broadcast messages.

Network Topology Simulation

Spring '17

Simulated a network topology using NS2 and NAM and analysed the logs using xgraph and AWK scripts.

Similairty Metrics for Catalysts

Spring '17

Visualized oxide catalysts based on various similarity metrics computed on the feature vectors representing their properties.

Fall '16

Performed hotspot analysis to identify performance bottlenecks in the given sequential software.
Improved performance by converting the sequential bottlenecks to parallel while maintaining synchronization using OpenMP.

MIPS Architecture Simulation

Fall '16

Implemented MIPS Single Cycle, Multi-cycle and Pipelined architectures in Verilog supporting various instructions.

CUDA Matrix Operations

Fall '16

Implemented various vector and 2D matrix operations in CUDA with performance optimizations.

Fastest Worder First

Fall '15

Multiplayer game in Node.js and Socket.io to make valid words starting with an assigned alphabet in one minute.

February '16

Application to automatically determine and focus on the dominant color in the selected box for an image while all other parts convert to greyscale. Implemented in C++ using OpenCV.

April '15

Made a fully responsive website for NFCFA '15 organized by Department of Chemistry. Made using Materialize library. The backend included sending registration confirmation emails using PHP.

March '15

Updates stock data from the web into the database periodically. Allows to query the data from mobile in JSON format using URL routes. Implemented in Python using Flask, SQLAlchemy and Shell.

Fall '14

A sticky notes Java application that recognizes text written in an image using Tesseract OCR engine and groups the notes based on hashtags in the recognized text.