Guided Labs

Build a tiny neural network without a training framework, then reproduce it with autodiff. Inspect exploding and vanishing gradients. Build a single transformer block, verify causal masking, and profile prefill versus decoding. Finally compare prompting, retrieval, and low-rank adaptation under one evaluation contract.

Evidence and review

Injected failures include a detached tensor, padding leakage, an incorrect mask, and evaluation examples copied into training. Claims must trace to versioned code and artifacts. Screenshots, unexplained generated prose, and a single favorable run do not pass.

Oral defense

Explain the mechanism, identify the most dangerous assumption, interpret one failure from raw evidence, and name the condition that would reverse your decision.

Source backbone

Use Dive into Deep Learning, PyTorch documentation, primary papers, and Designing Data-Intensive Applications selectively.