KAUST AI Intern

King Abdullah University of Science and Technology (KAUST)

July 2023 – August 2023

ROS and AI Intern

Smart Methods

June 2023 – August 2023

The Bias Within: Evaluating the Generation and Classification Dichotomy in Large Language Models

We conducted two bias detection experiments using GPT-3.5-generated data. The dataset consisted of both biased (with sociocultural stereotypes) and unbiased stories. GPT-3.5 achieved 53% accuracy in classifying its own biases, while the fine-tuned GPT-2 classifier achieved 97.5% validation accuracy. This shows the large performance gap between LLMs' generative and classification abilities. Highliting the need for task-specific training to develop high-quality bias assessments.

Additionally, we created a Flask interface for users to directly use the fine-tuned GPT-2 bias classifier.

Road Sign Prediction with GTSRB Dataset

This open-source project significantly improves the accuracy of CNNs for predicting road signs. Leveraging the GTSRB dataset, various enhancements to the CNN architecture are implemented, resulting in a notable increase in validation accuracy from 70% to 92%.

One key finding is that horizontal and vertical flipping, commonly used as augmentations, and sometimes as adversarial defense methods in literature, cannot be implemented for road signs due to label changes.

Numerical Methods for Complex Computations

Algebraic or "pen-and-paper" approaches may not always be sufficient for certain computations. Tasks like computing integrals, derivatives, or finding function roots can be challenging. This project focuses on exploring numerical solutions for such scenarios. Utilizing MATLAB for implementation.