The lab at night. The Professor smiles and points at a neural network on the screen, the Assistant types on a keyboard with a sleepy face, and the Fly Researcher in a helmet stands on the desk
Professor

Deep Learning A to Z from the Fruit Fly Lab

Welcome! Starting from the tiny brain of a fruit fly and ending at today's language models, we'll walk together through eight web pages. This isn't a book you just read. Press a button, and the pictures move and the numbers change.

Start with 3-1See the full list

The lab family

At the end of every part, the three of them each sum it up in a word or two.

The Professor, wearing glasses and a vest under a white coat
Professor

In charge of explaining. Untangles even hard formulas with a single picture. When excited, purple sparks fly from his hands.

The Assistant, with pink hair in two pigtails
Assistant

In charge of experiments. Runs code all night, and cries along with you when the results lose to the baseline.

The Fly Researcher, wearing a helmet and a white coat
Fly Researcher

The only one in the family with a real brain. Always has something to say when told that people and AI imitate its brain.

Research map — the eight parts in order

Read from the top down and the story flows on. It's also fine to pick whichever part interests you first. Each part lists "Recommended before this" at the very top.

How AI works and its partsTurning data into numbersOrder and attention

  1. 3-1

    How AI Works

    If people set the rules, it's machine learning; if it finds them in the data on its own, it's deep learning. We compare them side by side with the human brain and the fruit fly brain.

    Dim the lighting and see which method holds up.

  2. 3-2

    Building Blocks of Neural Networks

    Neurons, weights, activation functions, backpropagation. These are the parts you snap together one by one, like LEGO, to build a neural network.

    Follow the gradient down the mountain and shrink the error.

  3. 3-3

    Images and CNNs

    A photo is a grid of numbers. Scan it with a small magnifying glass (a filter) to find patterns, and it still recognizes the picture even when it's shifted a little.

    Shift and shuffle the digit 7 and pit an MLP against a CNN.

  4. 3-4

    Speech

    Sound is vibration in the air. Measure the vibration, cut it into pieces, split it by pitch, and it becomes a picture made of numbers.

    Listen to synthesized sounds yourself and compare 3-bit with 16-bit.

  5. 3-5

    Text

    Characters become numbers, numbers become pieces (tokens), and pieces become coordinates that carry meaning (embeddings).

    Watch "fruit fly" and "mosquito" move closer together without anyone teaching it.

  6. 3-6

    Time Series & Stock Prices

    First you have to beat the simplest answer: "tomorrow will be the same as today." Try it with real stock prices and it's harder than you think.

    Check four pitfalls with numbers. This is an educational example, not investment advice.

  7. 3-7

    Sequence Models

    Neural networks that walk one step at a time, carrying a memory. See what RNN, LSTM and GRU remember, and for how long.

    Flip the first signal and see whether the memory 20 steps later flips with it.

  8. 3-8

    Attention & Transformers

    Instead of relying on memory, it looks back at what came before whenever it needs to. This is the backbone of today's language models.

    Calculate Q·K·V by hand using a library analogy.

A day in the lab

You'll meet these scenes again at the end of each part.

The Professor holds a pointer in front of a blackboard full of equations
The Professor at the blackboard — 3-2, 3-3
In a room with whirring server fans, the Assistant holds up a keyboard and shouts
When training fails yet again — 3-4, 3-7
The Professor debugs in front of an error screen at 3 a.m.
Debugging at 3 a.m. — 3-6
The Professor presents a neural network on stage, helped by the Assistant and the fruit fly
Conference talk — 3-8

How to get the most out of it

Just press buttons

Every picture can be changed with buttons, and the text below updates to match the result. There's nothing you need to type.

Underlined words

Put your finger or mouse on a term with a dotted underline to see a simple definition. Each part also has its own glossary.

Run it yourself in Colab

Download the notebook from the "Code" section of each part and upload it to Google Colab, and you can recompute the numbers on the page exactly as shown.

For instructors and tutors

Grading criteria for each part's check questions and sample answers for the Colab exercises are collected on the answer key page for educators.

Phones are fine too

It's built to work on small screens as well. For part 3-4 Speech, turn off your device's silent mode to hear sound.