Train
From-scratch trainer
Data, model, loop, checkpoints, plots you can defend.
Deep Learning Specialisation
From tensors to trained networks you can explain: architecture, optimisation, regularisation, and a model you train — not a notebook you downloaded.
Created by Baljeet Dogra
Course Objectives
- Implement a network in PyTorch and know what every tensor is doing.
- Diagnose exploding loss, underfitting and leakage.
- Choose an architecture because of the data, not a tweet.
- Report metrics on a held-out set you did not tune against.
What you will produce
Report
Eval write-up
Held-out metrics, ablations, and what you would try next.
8 weeks curriculum
Expand a part for the syllabus. Content stays searchable when closed.
01 Tensors and autodiff Weeks 1–2
Shapes, broadcasting, backprop as implemented, not as a meme.
- Shapes
- Autograd
- Devices
- Numerical issues
02 Architectures Weeks 3–4
MLPs, CNNs, sequence models, a working-depth look at attention.
- MLPs
- CNNs
- RNNs
- Attention
03 Optimisation and regularisation Weeks 5–6
SGD, Adam, schedules, dropout, augmentation, early stopping.
- Optimisers
- Schedules
- Reg
- Generalisation
04 Train and defend Weeks 7–8
A real dataset, leakage checks, ablations. Capstone report.
- Data
- Leakage
- Ablations
- Capstone
Who this is for
Engineers past sklearn
You can fit a booster. Networks are still a black box.
Applied ML practitioners
You fine-tune APIs. You want to own a training loop.
Prerequisites
- Python and NumPy
- Mathematics for AI or equivalent: gradients, probability
- A GPU or a hosted equivalent for later weeks
Not a fit if
- A first Python course
- People who only want to call an LLM API
Related: Mathematics for AI · How to Become an AI Engineer
Questions
PyTorch or TensorFlow?
PyTorch. The ideas transfer.
Is this pre-training GPT from scratch?
No. You will understand the pieces. Full pre-training is Custom LLM Training.
Ready to start?
Eight weeks to a trained network you can explain. Create an account to enrol.
Enrol now