Gradient methods for classification

The problem

Optimization performance depends on the dataset, step-size strategy, and block-selection rule. Comparing methods required examining both accuracy and computational cost.

My contribution

• Co-authored an academic project comparing full gradient descent and block-coordinate methods.
• Explored Gauss–Southwell and randomized selection with different step-size strategies.
• Used synthetic data and MNIST experiments in Python with NumPy and PyTorch.

Evidence & scope

The public repository documents datasets, algorithms, accuracy, and CPU-time comparisons. Results are specific to the experimental setup.

What came out of it

Co-authored an implementation and comparative report on optimization methods for classification.

Produced a comparative research implementation and report with the project team, rather than a universal performance claim.