Using Label Word Embeddings in Image Classification
Machine Learning and Data Mining Course
Project on using label semantics for Image Classification to train models robust to adversarial examples. Rather than using cross-entropy loss, loss function using word embedding-based representation of labels was devised with significant results on Fashion MNIST and CIFAR-10 datasets using shallow networks.
Multithreading vs Python GIL: A study
Parallel Computer Architecture and Programming
Project involved studying the effects of the Python Global Interpreter Lock on libraries commonly used for machine learning (NumPy and Scikit-learn) by profiling for concurrency, locks, hotspots of time etc. An adaptive way to set GIL check interval was introduced as a result of profiling.