A two-hour, self-paced session for working developers

Deep Learning,
from the ground up.

Four architectures — ANN, CNN, RNN, LSTM — explained from neuron to gradient, with animations you can scrub, equations kept light, and code in Keras.

~120minutes
28sections
4architectures
8live animations
Section 01 · 5 min

Orientation: what we mean by "deep learning"

Before we touch a neuron, let's draw a map. Where does deep learning sit, and why do four architectures keep showing up?

Classical programming asks you to write the rules. Machine learning flips it: you show the machine examples and it infers rules from data. Deep learning is a particular shape of machine learning, where the model is a stack of differentiable layers, trained end-to-end with gradient descent.

A field map. The four architectures we'll cover are differentiated by what kind of structure they assume in the data.

If you remember only one thing from this session, make it this: each architecture exists because someone noticed a different kind of structure in real-world data, and built that assumption directly into the model.

ANN
No structure assumed
Tabular data, hand-crafted features. Every input is "just a number." The base case.
CNN
Spatial locality
Pixels near each other matter together. Same edge means the same thing wherever it appears.
RNN
Temporal order
"Bank of the river" ≠ "river bank account." Order is information.
LSTM
Long-range dependency
An RNN that doesn't forget after 8 tokens. A pacemaker for memory.

We'll keep math light: equations show up only where they earn their keep, always paired with plain English. Code blocks are TensorFlow / Keras. Skim them on first pass; come back later when you build.

~120 min remaining