Shikhar Ahuja

Georgia Tech Computer Science Student, passionate about creativity through machine learning, app development, chess and music
Scroll down to read about my experiences and projects

See Details

Machine Learning Projects

Paper on Novel Activation Functions

Published on arXiv

The Rectified Linear Unit is currently a state-of-the-art activation function in deep convolutional neural networks. To combat ReLU’s dying neuron problem, we propose the Parametric Variational Linear Unit (PVLU), which adds a sinusoidal function with trainable coefficients to ReLU. Along with introducing nonlinearity and non-zero gradients across the entire real domain, PVLU acts as a mechanism of fine-tuning when implemented in the context of transfer learning.

On a simple, non- transfer sequential CNN, PVLU substitution allowed for relative error decreases of 16.3% and 11.3% (without and with data augmentation) on CIFAR-100.

PVLU is also tested on transfer learning models. The VGG-16 and VGG-19 models experience relative error reductions of 9.5% and 10.7% on CIFAR-10, respectively, after the substitution of ReLU with PVLU.

When training on Gaussian-filtered CIFAR-10 images, similar improvements are noted for the VGG models. Most notably, fine-tuning using PVLU allows for relative error reductions up to and exceeding 10% for near state-of-the-art residual neural network architectures on the CIFAR datasets.

View Paper on arXiv.org

Research in Computer Vision

University of Central Florida

Since Jul '21, I have been working in the Computer Vision group under Professor Abhijit Mahalanobis

My responsibilities included analyzing research papers and creating neural net models for image recognition. I helped set up Convolutional Neural Nets for small and low resolution images with 98% accuracy. I also implemented 3-D point nets using Kitti Dataset and helped with data visualization.

Special thanks to Dr. Mahalanobis, Jibanul Haque, and Irene Tanner for their help and guidance.

Modeling Football Wide Receiver Performance

Individual Research

I used college football data (NCAA) to build a machine learning model and use the model to predict success of rookie Wide Receivers in professional NFL games. The goal was to determine which attributes can be used to accurately model a player’s performance

I obtained data on NCAA and NFL data from multiple sources – PFF, NextGen, ProFootball. I implemented a Linear Regression model using ols and sklearn Python libraries, measuring the success of the model using the minimum least squares number. I was able to achieve a 88% adjusted r^2 accuracy of the linear regression model

I predicted the NFL yards based on the model built using their college performance. The main observation were:

1. Attributes obtained from “combine” data increased the accuracy of the prediction. This included data such as the 40 yard dash, 3 Shuttle Drill, Broad Jump, etc.

2. By separating data into conference (groups of colleges) and modeling each one, I was able to predict 8 out 12 players correctly


My Mobile Apps

During the pandemic lockdown chess tournaments were canceled, and I wanted to find a way to reach chess students at various levels and share the key principles that I have learned over the last decade. I decided to create an app which provided interactive games and exercises to reiterate these key principles in a fun and engaging way. As I progressed, I added a lot of features which could be useful for chess students. I developed and published two apps: Chess Prof: Learn by Principle and Chess Prof: Collaborate and they are described below.

Chess Prof: Learn By Principle App

Overview

Chess Prof: Learn by Principle is a coaching app. It lays out key principles of chess along with exercises that reinforce the principles. These are organized by different levels of proficiency.

Beyond the coaching exercises, the app allows capturing chess games for analysis with full branching/annotation capability, making it a handy capture tool to use during tournaments. Users can also email the games in PGN notation.

The app provides the ability to submit their game to ChessProf for analysis, and to view the annotation once I respond.

Technical Info

I used Flutter for developing the app, as it allows deploying into both Apple and Android platforms. For storing data I used Google Cloud Firestore which allows streaming of update real time into mobile devices. I encountered several technical challenges which you can read about by clicking below.

Read More

Key Features

Principles by Skill Level

The app lays out key chess principles for Beginner, Novice, Intermediate and Advanced Levels

Interactive Exercises

The user can play the exercise games move by move, and view comments. Occasionally, there is a popup reiterating the principles, or asking the user to guess the next move.

Analyze, Save and Email Games with Variations/Annotations

The app allows capturing games with branching, merging, undo, delete, and comments. This can be emailed as a standard PGN file which be uploaded into any other chess software.

Free Expert Analysis

Users can submit their games for analysis, and see the comments by move once I send back comments

Login using mobile device, Apple, Google, User/Password

In order to save games, users need to create an account and have a choice of login mechanisms

Chess Prof: Collaborate App

Overview

Chess Prof: Collaborate, was developed for coaches and students to collaborate, organize and share games. This app was inspired by Grandmaster Magesh Panchanathan, who runs a chess club called Chess Kings and Queens, and wanted a way to see his students’ games organized into folders and give real time comments to them while they were in a tournament

Technical Info

This app was built as an extension to the Learn by Principle app, but I was able to add some cool features. For example users can check on "stay signed in", avoiding the need to login each time. Also I implemented email capability with this app, and subsequently added that feature in the Learn By Principle app as well.

Read More

Key Features

Organizing games into folders

The app allows a full set of create, update, delete operations for organizing games into folders. Users can also select a default folder (such as an ongoing tournament) for ease of navigation.

Coach - Student Interactions

Students can submit games to their coaches for review. As coaches add comments, students can see them in real-time. Once completed a coach can inform the student to save the "annotated" game for the future.

Flexible Lookups

Coaches can view a student's games either by folder, or just the ones submitted for analysis, or all games. Similarly students can view games by folder, the ones with coach's comments, or all of them.

Verification and Flexible Access Methods

Coaches can verify students in order to receive mail - this avoids them being spammed. Apple, Google, Mobile Device or Username/Password can be used to register and login, and users can "Stay Signed In" for ease of use

Chess Prof Logo

I had a lot of fun creating a logo for my Chess Prof app. I wanted to show a chess piece with a graduation hat, and this is what I came up with. The pictures below show my original drawing and how I digitally enanced it to get the final logo. The background of the logo is also used in the chessboard mosaic on the home page of the app.

Latest News Latest News Latest News

 

oxyGEN Scholarship

I won the National oxyGEN Scholarship for the best STEM video submission. I talked about the magic of the number 2, using examples from fables and computing to illustrate how exponentiation can be harnessed.

The magic of the number 2













Musical Projects

I've been dabbling with recording and mixing multiple Euphonium tracks to create covers of star wars themes. I am planning on creating an album of Euphonium covers for a variety of genres.

I really enjoyed arranging the tracks to bring out the harmony while using the Euphonium's power. An important step was the orchestration, where I figured out which harmonies to play by listening to the original piece. I wanted to maintain the style and sound of the original music, whilst giving my own flavor.

Star Wars Force Theme for three Euphoniums













Princess Leia's Theme for three Euphoniums