Projects
Ph.D. Projects - Summary
1. Effect of Sampling on Robustness of LIME
LIME(Local Interpretable Model- Agnostic Explanations) suffers from lack of robustness
Presented a modification of LIME based on CTGAN sampling and analysed it on a popular adversarial attack.
Demonstrated better robustness of the algorithm by comparing its explanations to vanilla LIME’s and calculating the percentage of time it could recognize adversaries
2. Clustering Guided GPUCB
Presented a geometric view of GP-UCB based on posterior mean and variance
Introduced two clustering-based acquisition functions GPUCB_NN & GPUCB^2
Reduced the search space of the GP-UCB acquisition function to a single best cluster chosen by these methods
3. Phishing Detection using Classical ML & ensemble of classical Quantam ML models
Developed an end-to-end system using which, a person can identify if the link received is malicious or not, just by entering the URL on our website.
To pre-process the data tokenization, stemming and vectorization have been utilized so that it can be used by algorithms.
Ensembled a Quantum Machine learning algorithm with LGBM, also used four classical machine learning algorithms, one bagging based and one boosting-based algorithm.
Obtained the highest test accuracy of 96.5% using XG Boost & Logistic Regression Classifiers
M.Tech Projects - Summary
- Lithography Hotspot Detection
Explored ICCAD 2012 benchmark dataset and used it to detect lithography hotspots using Convolutional Neural Networks (CNNs), Artificial Neural Networks (ANNs) and Vision Transformer (ViT)
ViTs gave an overall accuracy of 98.05% which is 1.39% higher than accuracy given by CNNs and 2.04% better than accuracy of ANNs
While comparing to other works we found that ViTs prove the best in terms of overall accuracy, but at dataset level its performance can be improved for two out of five datasets
2. Empirical Laws of Natural Language Processing for Neural Language Generated Text
Pre-processed and explored data, tuned hyperparameters to generate text using Long Short Term Memory Networks (LSTMs) and Generative Pretrained Transformer-2 (GPT-2)
Results showed that the text by LSTMs and GPT-2 followed Zipf’s law and Heap’s law
The LSTM generated text improved as the value of hyper-parameter Temperature increased
The comparison between GPT-2 and LSTM generated text showed that text generated using GPT-2 is more grammatically accurate than that generated by LSTMs
3. Analysis and Classification of Bike Share Data
Compared bike-share system usage between three cities in the USA by analyzing various descriptive statistics using Python, applied Min-Max Normalization and One-hot encoding to pre-process the data
Optimized different supervised learners for predicting whether a given person will buy subscription or not and obtained highest accuracy (.83) and F1 score (.78) using SVM