Not to complete the phase. Every core exercise is sized for CPU or a free hosted notebook GPU, and the capstone has a scaled-down path. Where a bigger run would change the conclusion, that is stated rather than glossed over.

Is backpropagation by hand really necessary in the autograd era?

For one week, yes. Autograd removes the arithmetic, not the need to understand what a vanishing gradient is or why a residual connection helps. People who have never done the backward pass tend to treat training failures as luck.

Where does this leave off relative to the GPT sprint?

The Build a Small GPT sprint is roughly weeks nine to eleven of this phase, taken alone. If transformers are the only part you want, the sprint is the cheaper route; this phase is what makes the transformer feel like a consequence of what came before it.

Phase courses₹1,499

Deep learning from a neuron up

Take this if you want to build and train modern architectures yourself rather than fine-tune somebody else's from a notebook.

What you’ll be able to do

You can derive and implement backpropagation, train networks that actually converge, build CNNs and transformers from their components, and fine-tune a pretrained model efficiently.

What it assumes

The Foundations phase or equivalent maths, plus comfort implementing models in Python. Classical ML helps but is not required.

Price: ₹1,499
Length: 16 weeks
Sessions: 16
Effort: ~8 h/week · 128 h total
For: Class 11–12 · College · Working professional

Adds it to My courses on your dashboard so you can find it again. Payment is not open yet — nothing is charged.

What’s taught, in order

01
Networks and backpropagation, by hand
A neuron, a layer, then the backward pass derived and coded with no framework. Everything afterwards rests on having done this once without autodiff to hide behind.
02
PyTorch, and training that converges
Tensors, autograd and modules, then the real business of training: initialisation, normalisation, learning rate schedules and the diagnosis of a loss curve that has gone wrong.
03
Convolutional networks
Convolutions, pooling and the architectures that shaped the field, with an emphasis on reading a paper's architecture diagram and reproducing it.
04
Sequences, then attention
Recurrent networks and their limits, which is what makes attention feel inevitable rather than arbitrary when it arrives in the following week.
05
Transformers and tokenization
Building a transformer end to end, then tokenization in the depth it deserves — the layer where a surprising share of real model behaviour is decided.
06
Pretraining, fine-tuning and PEFT
What pretraining does, how transfer learning exploits it, and parameter-efficient methods that let you fine-tune a large model on hardware you own. Ends in a capstone.

What you build

Choose a real task with unstructured data and carry it end to end: establish a baseline, implement the core architecture yourself rather than importing it wholesale, train it to convergence with your diagnostics documented, then fine-tune a pretrained model on the same task and account honestly for the difference in results and in compute.

Check it yourself against this

The core architecture is implemented by you and verified against a reference on a small input.
Loss curves are shown for at least one failed run, with the cause diagnosed.
Your from-scratch model and your fine-tuned model are compared on the same test set.
The compute cost of each approach is reported, not just the accuracy.
One deliberate ablation is run, and its result is interpreted.
You state which result you would actually use and why.

Then answer this

Your fine-tuned model almost certainly beat your from-scratch one — what did the pretraining supply that sixteen weeks of your own effort could not?

Two or three sentences, in your own words. If you cannot, the course is not finished — go back to the module it came from.

What this course does not do

Training runs are sized for a laptop or a free hosted GPU, so you will not reproduce a frontier result and the course says so. Reinforcement learning, graph networks and speech are out of scope; generative modelling gets its own phase.

Where this leads

The Deep Learning phase page breaks all sixteen weeks out individually, including the optional computer vision specialisation week, so you can see where attention arrives and what precedes it.

See the sixteen-week plan Try a real lesson free

FAQ

Do I need a GPU?: Not to complete the phase. Every core exercise is sized for CPU or a free hosted notebook GPU, and the capstone has a scaled-down path. Where a bigger run would change the conclusion, that is stated rather than glossed over.
Is backpropagation by hand really necessary in the autograd era?: For one week, yes. Autograd removes the arithmetic, not the need to understand what a vanishing gradient is or why a residual connection helps. People who have never done the backward pass tend to treat training failures as luck.
Where does this leave off relative to the GPT sprint?: The Build a Small GPT sprint is roughly weeks nine to eleven of this phase, taken alone. If transformers are the only part you want, the sprint is the cheaper route; this phase is what makes the transformer feel like a consequence of what came before it.

Other courses

This is a standalone course. It does not add weeks to the 78-week curriculum or change your roadmap. If you want the whole path instead, the full programme is ₹399/month · ₹2,000/year — see what it covers.

Deep learning from a neuron up

What you’ll be able to do

What it assumes

What’s taught, in order

Networks and backpropagation, by hand

PyTorch, and training that converges

Convolutional networks

Sequences, then attention

Transformers and tokenization

Pretraining, fine-tuning and PEFT