Poornika Bonam poornikabonam

🚀Hello there! Welcome to my profile 🌌

Master of Science in Computer Science, University of Massachusetts Lowell (Dec 2023)
Bachelor of Technology in Computer Science & Engineering, Jawaharlal Nehru Technological University, Hyderabad, India (Aug 2021)

Developed custom-built LSTM models for text generation to ensure semantic consistency.
Integrated additional contextual information using Word2Vec embeddings and K-Means clustering techniques.
Conducted thorough preprocessing of the dataset to eliminate noise and tokenize the text effectively.
Implemented and trained the models using Python programming language.
Achieved notable advancements in both relevance and coherence of the generated text, showcasing the potential of context-aware LSTM models in advancing natural language processing capabilities.

Implemented a Deep Averaging Network (DAN) in PyTorch for quiz bowl questions, achieving 85.25% accuracy in category prediction.
Developed a Deep Averaging Network (DAN) model for answer prediction, starting with question analysis using spaCy for entity extraction and TF-IDF score for relevant Wikipedia page retrieval.
Constructed the DAN model from scratch, incorporating linear layers for classification, without relying on external libraries.
Implemented training functionality using the Adamax optimizer and Cross Entropy Loss, with careful consideration of model optimization techniques.
Enhanced model performance by initializing word embeddings with GloVe vectors, resulting in a notable 3% increase in accuracy on the test set.
Addressed challenges such as lengthy training times due to content retrieval, demonstrating the robustness and adaptability of the developed model.

Explored vulnerabilities in deep learning models through adversarial attacks, specifically focusing on FGSM attacks on the MNIST dataset.
Investigated the efficacy of various defense mechanisms including Adversarial Training, Adversarial Training with Gradient Masking, Label Smoothing, Distilled Neural Network, GDA with ReLu, and GDA with BReLu.
Implemented and evaluated each defense mechanism's performance against adversarial attacks using TensorFlow, Matplotlib, and NumPy for analysis and visualization.
Recorded the accuracy of each defense mechanism against FGSM attacks on MNIST, with results ranging from 1.82% to 90.50% accuracy.
Presented findings indicating the effectiveness of certain defense mechanisms, such as Adversarial Training combined with Distilled Neural Network, which achieved a notable accuracy of 90.50% against FGSM attacks, demonstrating promising defense strategies for mitigating adversarial vulnerabilities in deep learning models.

Implemented BERT and Naive Bayes for sentiment analysis, focusing on hate speech detection in Twitter data related to major incidents.
Explored preprocessing techniques like Lemmatization and stemming to standardize tweet content and employed feature extraction methods such as TFIDF and bag-of-words.
Balanced the dataset containing a high proportion of positive tweets using class weights and downsampling to improve model performance.
Evaluated model effectiveness, with the Decision Tree Classifier achieving the highest accuracy of 86% using TFIDF and VowpalWabbit reaching 88% accuracy with an average loss of 12%.
Identified future research directions including expanding the dataset with more hate tweets and enhancing model training to improve classification accuracy in combating societal issues.

The project targets the pervasive issue of fake news dissemination, employing a novel approach integrating CNN and LSTM architectures.
Methodologies encompass the adoption of dimensionality reduction techniques like PCA and Chi-Square, enhancing feature extraction efficacy within the neural network framework.
Experimental validations conducted on the FNC dataset demonstrate notable improvements, with PCA outperforming Chi-Square and achieving a 4% accuracy boost.
Implemented using Python, the project provides a practical solution for automated fake news detection, potentially mitigating the societal impacts of misinformation.
Significantly, the project's focus on stance detection and determining news article credibility vis-a-vis headlines contributes to advancing automated fake news detection tools, addressing critical needs in contemporary media ecosystems.

The project involves implementing a home surveillance system using Raspberry Pi and Python.
It aims to detect motion, capture images, and send them to Dropbox for storage and notification.
Leveraging Raspberry Pi's capabilities along with Python and OpenCV, the system detects motion through image processing techniques.
Integration with Dropbox provides a reliable method for storing captured images and enabling remote access.
The primary objective is to develop a smart surveillance system to ensure home safety by detecting and alerting users of any unusual activity.