Modern Deep Learning and Representation

Part of the Applied AI Security and Assurance curriculum, authored by Cameron Hopkin, CISSP, CEH, CHFI. Published as a reference, not an enrollable course.

Description

Replaces the dated bag-of-words and GAN-centric approach with the actual modern stack: transformers, the Hugging Face and PyTorch ecosystem, and diffusion as the generative flagship. The models that define the field are built in this ecosystem [5].

Outcomes

• Explain and implement the transformer end to end.
• Fine-tune open models with LoRA and PEFT.
• Reason about what representations encode and leak.

Modules

1. Transformer architecture, attention, tokenization, and embeddings.
2. Training dynamics and transfer learning.
3. Fine-tuning, LoRA, and PEFT.
4. Diffusion models as the modern generative flagship, with GANs treated as history.
5. Interpretability primitives and probing.
6. Representation leakage.

Signature lab

Fine-tune an open model with LoRA using the Hugging Face ecosystem and document precisely what changed in its behavior and why.

Research thread

What a model's internal representations reveal, and what that leaks to an adversary.