Selected Projects

• Project of Coursework "Deep Learning

• Problem Statement: This project explores multiple deep learning architectures (ResNet-18/50, VGG-13/16) and data augmentation techniques (CutMix, Random Erasing, AutoAugment) on Tiny ImageNet, revealing that CutMix consistently offers the best training efficiency and test accuracy. Curriculum Learning, particularly Local-to-Global Learning (LGL), outperforms vanilla training, though combining LGL and CutMix does not yield stronger generalization. Attempts to use Conditional Diffusion Models for augmentation face reconstruction challenges, highlights the superior benefits of data augmentation and curriculum learning for enhanced classification performance.

• Skills Acquired: Deep Learning,  Data Augmentation, Curriculum Learning, Probabilistic Diffusion Models

• Project Details: https://github.com/shilpa2301/Data-Augmentation-and-Curriculum-Learning-on-TinyImageNet 

• Master's Thesis under Dr. Guni Sharon, Texas A&M University : Link

• Problem Statement: A recent study introduced Gran Turismo (GT) Sophy, an AI racing agent trained using Deep Reinforcement Learning (RL) to outperform top human drivers in the PlayStation simulator Gran Turismo Sport (GTS), with skills in race-car control, racing tactics, and etiquette. This research explores adapting GT Sophy’s race-car control algorithm for real-world small-scale cars, addressing two key gaps: reliance on high-level simulator data (inputs, rewards, and static maps) and the impractical training demands of extensive time and computational resources. To overcome these challenges, the study proposes three solutions: (1) using camera and IMU sensor data with State Representation Learning (SRL) to reduce state-space complexity, (2) designing rewards based on IMU data to improve stability, and (3) implementing action limit clipping to minimize lap times. These solutions were tested on the Donkey simulator and a JetRacer RC car, showing that SRL, combined with a clipped throttle range and stability-focused rewards, significantly improved sample efficiency for completing tracks and reducing lap times, paving the way for real-world applications. 

• Skills Acquired: Deep Reinforcement Learning,  Sim-to-Real

• Project Details: Since the implementation was not made public by the authors, I implemented the QRSAC algorithm.

https://github.com/shilpa2301/QRSAC

• Project of Coursework "Robot Spatial Intelligence and Navigation"

• Problem Statement: As a part of the project, I created a robot from scratch in ROS2 , created its LiDAR sensor, its maps and finally used Bi-directional RRT to create dynamic planning. With a user prompt, the agent has the capability to change its planning online.

• Skills Acquired: Planning and Navigation, ROS2

• Project Details: https://github.com/shilpa2301/Robot-Simulator-creation-and-Navigation

• Secured 1st Position in the competition (Certificate)

• Problem Statement: LibAUC library was overfitting smaller dataset for AUCMLoss. We had to optimize AUCMLoss and compare with 7 datasets, namely BreastMNIST, PneumoniaMNIST, ChestMNIST, NoduleMNIST3D, AdrenalMNIST3D, VesselMNIST3D, SynapseMNIST3D. Among these seven datasets, ChestMNIST is a multi-label classification tasks, and others are binary classification tasks. For ChestMNIST, each label is considered as a binary classification problems. Except for ChestMNIST, other datasets are relatively small. Your goal is to improve the benchmark performance reported in the MedMNIST paper. For fair comparison, we had to use the same network structure as in the MedMNIST paper. We have used Resnet-18.

Goal: This project aims to improve the generalization ability of DAM for medical image classification tasks.

Collaborator: Anushka Garg, MS CS Student, TAMU

• Skills Acquired: Deep Learning,  Deep AUC Maximization (DAM)

• Project Details: 

https://github.com/shilpa2301/Deep-AUC-Maximization 

• This is a research project conducted as a part of coursework. The work estimates the relative performance of various combinations of group recommendation techniques when compared to one another. Based on the above comparisons, we developed an optimal pipeline for group recommendation techniques, based on efficiency or performance. In addition to this, we also figured out the factors/techniques contributing more to optimality and the ones contributing less or deviating from or hindering optimality

• Skills Acquired: Recommender Systems, Group Recommendation

• Project Website: 

https://sites.google.com/tamu.edu/group-recommendation 

• The project includes processing of road landmarks and other relevant data and storing the data in upstream for creating HD map for autonomous driving. In the downlink, it fetches these stored data from map and accurately localizes the vehicle by improving and predicting (based on need) GPS location of the vehicle. Finally, the data needs to be given out in an intelligent manner to fit the available bandwidth. The project is implemented completely in C++ and ported to Renesas boards to work on them. Worked on Module testing and unit testing, ensuring MISRA Quality standards for each development activity as well. Also need to adapt requirements based on camera generation update or customer requirements.

• Skills Acquired: Automated Driving Assistance System (ADAS), Computer Vision, Machine Learning, Geodesy and Coordinate Systems (Geodetic, ENU, NED, ECEF), Vehicle State Estimation and Position Localization.

• The project includes detection and tracking of a particular kind of static object (wireless charging pad) on ground from a designated far-away distance using object detection mechanism. I had to develop and implement the algorithm for detection and recognition of the object using various Computer Vision and Machine Learning algorithms and compared their performances. Languages used were Python, MATLAB, C++ for various experiments of comparison.

• Skills Acquired: Computer Vision, Machine Learning, Object Detection, Object Recognition, Static Object Tracking

• NOMINATED as the only project from B.Tech. CSE Department for GOLD MEDAL (2017) IN THE INSTITUTE: Chosen two particular fields of Digital Forensics: Copy-Move Forgery in Videos and Video Source Identification. Copy-Move Forgery in Videos involve inserting or replacing a set of frames from another part of the same video.  After Literature survey, an existing work has been implemented. One of the drawbacks is chosen to be addressed is video forgery involving insert or replace of a subsequence of frames from a different video. The algorithm has been proposed and implemented. The problem gave rise to another problem- data origin authentication in videos from multiple sources. This problem has also been completed with satisfactory detection accuracy.

• Skills Acquired: Copy-Move Forgery, Digital Forensics, Image & Video Processing, Source Identification, Machine Learning

• The project involved tracing the system calls of various malwares and benign executables along with their parameters and return values. The aim was to find any interesting pattern in the system-call executions those occur frequently and are common to all malicious programs, but not in benign executables. The frequent itemset data mining algorithms were used to find the patterns.

• Skills Acquired: Data Mining