Ali Tehrani

Automatic optimization of GPU Kernels.

• Trained an Encoder-Decoder transformer model from scratch for code translation and generation. • Improved Mask Language Modeling, Denoising Auto Encoding, and Back Translation training objectives. • Proposed new training objective, AST entity recognition, to further improve code reasoning capability of LLMs. • Devised a robust program graph representation for performance optimization and program analysis. • Developed a novel embedding technique for numbers that solves the issue of any unknown number at inference. • Developed a GNN-based DL framework for optimizing HPC kernels on NUMA Systems statically.

• Developed and Trained Fully Connected Neural Network for predicting phenotype from genetic data. • Utilized Amazon SageMaker to train the network. • Applied Lottery Ticket Hypothesis (LTH) pruning technique to prune more than 53% of the model’s weights. • Utilized explainable AI (SHAP) to identify the critical features in the training data. • Optimized data preprocessing and achieved a 3.26x speedup in comparison to the original implementation.

• Provided course material for teaching MS Excel and Access to more than 70 undergrad students. • Automated course exams and homework through the Pearson MyLab IT platform. • Managed and led a team of 7 teaching assistants, defined and distributed tasks among them.

• Formulated Data Race bug detection as a deep learning-based classification problem. • Developed a Convolutional Neural Network for Abstract Syntax Trees. • Achieved up to 86% of data race bug detection. • Introduced a noising strategy to automatically inject errors in the training data during data preparation automatically.

Courses: CS 527: Concurrent Systems CS 519: Design and Analysis of Algorithms CS 574: Introduction to Machine Learning CS 573: Machine Learning CS 665: Advanced Topics in Software Engineering CS 531: Theory of Computation