Study notes for all 13 courses, written to match the real module order — with animated explainers, verified papers and videos, and the exam traps that catch people out.
▶ Fly through the whole certificateA one-minute motion-graphics overview of the whole certificate — what the thirteen courses build toward and why the order matters. Sound on.
Made with the same pipeline course 13 teaches: generation, assembly, iteration.
scikit-learn, the model lifecycle, and every classical algorithm the certificate assumes you know
The mechanics of a neural network — neurons, gradients, backprop, and the loss table you'll use for the rest of the certificate
Everything Sequential can't do — functional models, transfer learning, transformers, generative models and reinforcement learning
Tensors, autograd, and the five-step training loop that every remaining course in the certificate is built on
Initialisation, dropout, batch norm, momentum and CNNs — the techniques that make a deep network actually train
The same pipeline built twice — Keras and PyTorch, side by side — plus vision transformers. The comparison is the assessment
The pivot point — from training models to adapting foundation models. Tokenization, data quality, and the GPT/BERT/T5 map
How text becomes numbers — one-hot to embeddings to Word2Vec to seq2seq, and the bottleneck that produced attention
The most important course in the certificate — attention, positional encoding, masking, and what separates GPT from BERT
LoRA, QLoRA and PEFT — the most immediately employable material in the certificate
Alignment — instruction tuning, reward modelling, RLHF with PPO, and the DPO derivation that removed the reward model entirely
Prompting, agents, and retrieval — where the certificate becomes application engineering
The capstone — a QA bot over your own PDF, and the one artefact from this certificate you can show someone in 30 seconds