Language is where deep learning had its biggest breakthroughs. This module starts with the fundamentals of natural language processing: turning text into tokens, tagging linguistic features, extracting named entities, and representing words as vectors. It then moves to the transformer architecture, where self-attention lets a model weigh every word in a text against every other, and which underlies models like BERT and GPT.
You'll work with text using spaCy, build a small GPT-style model to see attention and transformer blocks up close, and use the Hugging Face ecosystem to load pretrained transformers, work with tokenizers, and deploy your own demos.
These courses are long, so treat them as a reference: watch the parts you need, then come back for the rest. When you're done, continue on to LLMs & Prompt Engineering.
What NLP is and how spaCy works: containers, linguistic annotations, named entity recognition, word vectors, pipelines, rule-based components, regular expressions, and an applied financial NER project.
Build a GPT-style model in Python and PyTorch: embeddings, matrix math and gradient descent, then the transformer architecture, self-attention, multi-head attention, positional encoding, and pretraining vs finetuning.
The Hugging Face ecosystem end to end: the Transformers library (BERT, GPT-2, DistilBERT), speech and audio pipelines, Stable Diffusion with Diffusers, multimodal models, Gradio demos, and deploying apps on Spaces.