Gradient Descent and Optimization projects for B.Tech students — real examples

What students actually build

• Training-run diagnosis — learning rate, convergence, overfitting — is part of every model-building capstone.
• LoRA fine-tunes of open-weight models are a named capstone deliverable.

Labs and infrastructure

• Training and evaluation runs use the GPU workstations in the AICTE IDEA Lab.
• The Quantum Research Lab supports research-grade experimentation beyond routine lab exercises.

How VSET teaches Gradient Descent and Optimization

Gradient descent iteratively moves model parameters down the slope of a loss function; variants such as SGD, momentum and Adam control how large and how noisy those steps are. Every trained model in modern AI comes out of some version of this loop. At VSET this maps to documented coursework depth inside B.Tech CSE (AI & ML).

• Optimisation is the engine underneath the machine learning and deep learning material published at learn.engineering.vips.edu.
• Learning-rate behaviour, convergence and loss landscapes are taught before the transformer content that depends on them.
• The same optimisation loop reappears in the fine-tuning material, where only adapter weights are updated.
• Delivered inside the B.Tech CSE (AI & ML) track, one of VSET's seven GGSIPU-affiliated B.Tech programmes.