Visualizing and Interpreting Transformer-based Vision Models
Applying Shapley-value based methods (FastSHAP) to Vision Transformers and Masked Autoencoders, comparing results to a classical ResNet for model interpretability.
Things I've built or am building.
Applying Shapley-value based methods (FastSHAP) to Vision Transformers and Masked Autoencoders, comparing results to a classical ResNet for model interpretability.
Automatic pixel-level classification of satellite imagery using U-Net and ResNet architectures, achieving 0.887 mean IoU on the Landcover.ai dataset.
A learning-based approach for predicting performance of multi-threaded benchmarks using execution parameters and machine learning models.
Exploring Twitter feeds and news articles as alternative data sources to predict closing stock prices of leading green energy companies.
Training ensemble learning and deep learning models to predict GPU kernel execution time using CUDA-Flux profiler data from PTX code.