Course

Deep Learning: from one neuron to a transformer

Name: Deep Learning: from one neuron to a transformer
Price: 2199 BDT

Build a GPT one part at a time: a neuron, a trained network, convolution, attention, a transformer block, and the steps that turn a raw model into an assistant. A tiny name generator trains live in your browser and grows up with you.

73 lessons7h 43mEnglish

What you'll learn

Compute a neuron’s output by hand and explain why stacking layers needs a non-linear activation
Show how two layers solve XOR, the pattern a single straight boundary cannot separate
Take a gradient step by hand, choose a learning rate, and recognise divergence on a loss curve
Trace backpropagation through a small graph and say which weight to change first
Train a character-level model, read its loss against a baseline, and keep the checkpoint before overfitting
Compute a convolution and a feature-map size, and use reconstruction error to flag an unusual image
Explain attention as a soft lookup of query against key, compute one row of weights, and say why scores are divided by √d
Name every part of a transformer block in order and turn a final vector into next-token probabilities
Measure what a tokenizer charges for Bangla against English, and what that means for context and price
Describe how pretraining, supervised fine-tuning, preference learning and verifiable rewards turn a raw model into an assistant
Recognise reward hacking, and compute GRPO group advantages from a set of right and wrong answers
Read a Hugging Face model card and choose a model for a Bangla or Banglish task
State with dates what current agents can and cannot do, and argue both sides of the AGI debate

Requirements

The Machine Learning course or equivalent: fitting a line, train and test splits, embeddings and next-token prediction from counts
Comfortable with vectors, a weighted sum and reading a graph
A slope as “how much the output changes per unit of input”; no calculus beyond that is assumed
No programming is required: code is read, never run

Course content

73 lessons · 7h 43m

One neuron, then many

See the finished tiny model invent names, then build its smallest part. Weigh and add inputs in one neuron, bend the result with an activation, stack neurons into a layer, and solve the XOR rule a single line never could.

0/8

The whole game: a model that invents names

A 3,115-number model trained on 1,405 names writes new ones. By chapter 5 you will have built every part of it.

Preview3m 2.

One neuron: weigh, add, decide

Multiply each input by a weight, add a bias, pass it through a gate. Rain 2 and holiday 1 give 0.2 with ReLU.

Preview6m

Activation: the bend that makes depth pay

ReLU blocks negatives, sigmoid squashes to a probability. Without a bend, stacked layers collapse into one straight line.

A layer is a matrix times a vector

Each row of weights is one neuron. [[1, 2], [-1, 1]] times [2, 1] gives [4, -1], row by row.

Two layers solve XOR

Uncle Rafi’s rain-or-holiday rule beat the perceptron. Two hidden ReLU neurons get all four days right.

Tania’s quick-fire round: neurons

Compute a neuron, name the gate, read a layer’s output and spot which network can draw XOR. Keep the streak.

Exam practice: one neuron, then many

Quiz questions with the trap behind each wrong answer, then a neuron and a two-layer network worked by hand.

10m

One neuron, then many: quick revision

The chapter on one screen, a forward pass in order, one broken calculation to fix, and a spaced flashcard deck.

How a network learns

Turn wrongness into one number, step every weight downhill, and pass blame backward node by node. Then train a real name generator live, watch it beat counting, catch it memorising, and stop it at the right moment.

0/11

Loss: how wrong, as one number

Predict −1 when the answer is 1.5 and the squared error is 6.25. Squaring punishes big misses and ignores the sign.

Preview6m 2.

One step downhill

The slope −6.75 says raise the weight. A 0.08 step moves it from −0.5 to 0.04 and the loss falls from 5.06 to 2.07.

Preview7m

The training loop

Predict, measure, find the slope, update, repeat. Seven steps at 0.08 reach a loss under 0.01; past 0.44 every step makes it worse.

Many examples, one loss

Average the error over every example and step both weight and bias. Five points settle near w 0.75, b 1.25, MSE 0.048.

Backpropagation: blame flows backward

Each node multiplies the blame it receives by its own local slope. w1 gets −1.69, the steepest, and one nudge cuts the loss 0.289 to 0.055.

The name generator learns

Loss starts at ln 27 = 3.30. Counting pairs reaches 2.53; the network beats it by step 300 and gibberish turns into names.

Memorising is not learning: knowing when to stop

On Shumi’s 100-name shortlist, validation bottoms out at step 125 and rises while training loss keeps falling. Keep that checkpoint.

Dropout and weight decay

Switch off random neurons, or charge rent for big weights. On a 30-point curve, dropout 0.2 cuts the validation gap from 0.27 to 0.20.

Tania’s quick-fire round: training

Read a loss curve, pick the learning rate, pass blame through one node, and name the PyTorch line that does it.

10.

Exam practice: how a network learns

Quiz questions with the trap behind each wrong answer, then a gradient step, a backward pass and a checkpoint to justify.

10m

11.

How a network learns: quick revision

The training loop on one screen, its steps in order, one broken gradient to fix, and a spaced flashcard deck.

Seeing and squeezing

Slide a small filter over a picture and it finds edges on its own. Squeeze a pattern through a two-number bottleneck, walk the space of codes, clean noise away, and flag the tin the network cannot rebuild.

0/9

A filter slides over the picture

Nine weights, nine pixels, one sum. The edge filter scores 6 where the stamp’s outline turns, and the same nine weights work everywhere.

Preview7m 2.

Feature maps, stride and padding

A 3 × 3 filter on a 5 × 5 image gives a 3 × 3 map. Stride 2 halves it; padding 1 keeps the size. Output = ⌊(n + 2p − k)/s⌋ + 1.

Preview7m

Squeeze it through a bottleneck

An autoencoder must rebuild 16 pixels from two numbers. The cross motif rebuilds perfectly; an all-dark tin loses half its pixels.

Walking the latent space

Pick two clues and the decoder draws a pattern it never saw. (0.5, 0.5) is a faint cross: codes in between mean images in between.

Cleaning noise away

Three flecks of ink on a stamp: the bottleneck has no room for them, so the rebuild comes back clean. Find the damage it cannot undo.

Can’t rebuild it? It’s unusual

Score each tin by how badly it rebuilds: 0, 0.125, 0.75, 0.1875. A threshold between 0.1875 and 0.75 flags only the square misprint.

Tania’s quick-fire round: vision

Compute a filter response, size a feature map, read a rebuild error and choose a threshold. Keep the streak.

Exam practice: seeing and squeezing

Quiz questions with the trap behind each wrong answer, then a convolution by hand and a misprint check to design and defend.

10m

Seeing and squeezing: quick revision

Filters and bottlenecks on one screen, the misprint check in order, one wrong map size to fix, and a spaced flashcard deck.

Memory and attention

A Banglish order ends in "na" and the meaning flips. See why a relay memory fades, then build attention piece by piece: routing, query, key and value, the square-root scale, many heads, no peeking ahead, and word order.

0/9

Memory that fades

Pass the clue along word by word: the RNN pull on it vanishes to 6% after five. The LSTM forget gate keeps 73%. Attention looks straight back.

Preview7m 2.

Attention: every word asks the others

Each word scores every other word and spreads 100% of its attention by softmax. "na" puts 80% on "pathaben", the verb it flips.

Preview7m

Query, key and value

A soft lookup table: the question “Kokhon lagbe?” matches the time key, and 66% of the answer comes from “tomorrow 8 am”.

Scaled dot-product attention

Match, scale by √d, weigh, mix. Dividing by √4 = 2 keeps long vectors from turning softmax into a hard pick.

Many heads, many questions

One head links "20ta" to "shingara", another links "na" to "pathaben". Silence a head and that link disappears from the mix.

Word order and no peeking ahead

Without a mask, "shingara" puts 52% of its attention on words not yet written. Positions tell “Rafi pays Shumi” from the reverse.

Tania’s quick-fire round: attention

Read a weight, softmax three scores, scale by √d, and say which row of the mask is closed. Keep the streak.

Exam practice: memory and attention

Quiz questions with the trap behind each wrong answer, then one attention row by hand and a memory design to compare.

10m

Memory and attention: quick revision

Attention on one screen, its four stages in order, one wrong softmax to fix, and a spaced flashcard deck.

Building a GPT

Turn text into tokens and meet the Bangla token tax, keep the signal steady with residuals and layer norm, assemble one transformer block, turn a vector into next-token odds, sample with temperature, and retrain the name generator with attention.

0/11

Text becomes tokens

Start from bytes and merge the most frequent pair, again and again. Twenty merges shrink a 274-byte corpus to 157 tokens.

Preview7m 2.

The Bangla token tax

The same greeting costs 9 tokens in English and 59 in Bangla script: 6.6 times the price, the context and the wait.

Preview7m

Keeping the signal: residuals and layer norm

Add each update to the stream instead of replacing it, then rescale to mean 0 and variance 1. [2, 4, 6, 8] becomes ±1.34 and ±0.45.

One transformer block

Widen, bend with GELU, narrow: that is the feed-forward layer. Then assemble Norm, attention, add, Norm, feed-forward, add.

From a vector to the next token

Dot the last vector with every token’s row to get logits, then softmax. After “kal 20ta shingara”, pathaben leads at 44%.

Generation and temperature

Pick a token, append it, predict again. At 0.3 “cha” after “ek cup” is 99%; at 2.2 coffee rises to 35%.

The name generator gets attention

Same names, same seed, eight letters of context. Attention beats counting by step 150 and reaches 2.38 where the network managed 2.45.

Tania’s quick-fire round: the GPT

Count tokens, normalise a vector, order a block, softmax some logits and name the Hugging Face line. Keep the streak.

Mission: what does a Bangla order cost?

Train the tokenizer on Corner Bakery’s English orders, then price one order three ways. After 20 merges Bangla costs 87 tokens to English’s 16.

10m

10.

Exam practice: building a GPT

Quiz questions with the trap behind each wrong answer, then a layer norm, a softmax and a token bill worked by hand.

10m

11.

Building a GPT: quick revision

The whole GPT on one screen, a block in order, one broken probability to fix, and a spaced flashcard deck.

Many languages, many senses

Translate a Banglish order with an encoder and a decoder, put Bangla and English words in one space, match menu photos to captions, see how models read long documents and grow huge but cheap, and choose a Bangla model from its card.

0/8

Encoder and decoder: translating an order

“duto shingara pathan” becomes “send two samosas”. One squeezed memory drops the “two”; attention to the source keeps every word.

Preview7m 2.

One space, many languages

A multilingual model places “kal” 0.64 from “tomorrow” and “na” beside “not”. Meaning, not spelling, sets the neighbours.

Preview6m

Pictures and captions in one space

Pull each photo toward its own caption and push wrong pairs apart. Ten pulls match tea, samosa and sweets: the idea behind CLIP.

Long context and mixture of experts

Attention links grow with the square of the context: 8 tokens make 64, a million make 10¹². Experts let a huge model run like a small one.

Bangla models on Hugging Face

BanglaBERT for understanding, BanglaT5 for translating, multilingual families for chat. Read the card, load the tokenizer, test on Banglish.

Tania’s quick-fire round: many senses

Encoder or decoder, which pair matches, how many attention links, which model card line matters. Keep the streak.

Exam practice: many languages, many senses

Quiz questions with the trap behind each wrong answer, then an attention cost to calculate and a model choice to justify.

10m

Many languages, many senses: quick revision

Translation, shared spaces, context cost and model choice on one screen, a translation in order, one error to fix, and a flashcard deck.

From raw model to assistant

A raw model continues text; an assistant answers. Fine-tune on demonstrations, learn from preferences and watch a reward get hacked, reward only checkable answers with GRPO, see thinking longer emerge, and meet tools, agents and small models.

0/10

A base model is not an assistant

Pretraining teaches what text looks like. Fine-tune the name generator on town names and the same weights swing 0.24 nats toward them in 50 steps.

Preview7m 2.

Learning from demonstrations

Show the model one good answer and raise its probability. Supervised fine-tuning lifts P(correct) from 26% to 82% in five updates.

Preview7m

Learning from preferences, and reward hacking

A reward model that likes politeness scores a padded wrong answer 1.56. After ten updates the engine flags the hack: reward up, correctness down.

Rewards you can check: GRPO

Sample eight answers, mark each right or wrong, and score each against the group mean. Three right of eight: +0.625 and −0.375.

Thinking longer: reasoning and test-time compute

Reward only correct answers and the worked answer wins: P(correct) 0.905 after 20 updates, average length up from 12.6 to 22.1 tokens.

Tools, agents and small models

A model that writes a tool call, reads the result and loops is an agent. Distillation and LoRA make a model small enough for a laptop.

Tania’s quick-fire round: alignment

Name the training stage, compute a group advantage, spot the hacked reward and pick the TRL trainer. Keep the streak.

Exam practice: from raw model to assistant

Quiz questions with the trap behind each wrong answer, then group advantages by hand and a training plan to evaluate.

10m

From raw model to assistant: quick revision

The pipeline on one screen, the stages in order, one mislabelled reward to fix, and a spaced flashcard deck.

10.

What happens when a model acts in the world?

Every stage you built, then the dated evidence on agents and AGI, the definitions that disagree, the sceptics, and the course that comes next.

About the creator

Classytic