How can a computer tell a cat from a dog?
No two photos line up perfectly. A fixed rule would need to list every angle, light and pose.
Back to the lesson shelf
Neural
LESSON 01 The first layer
Neural
network.
A pile of small filters that pass a signal forward, one tiny decision at a time.
In one lineMany tiny votes → one useful answer.
NEURAL_NETWORK / 01 flowing
inputanswer
data → patterns → decision9 nodes / 17 links
Short version: data → filters → answer.Scroll, click, repeat.
00 Overview / start here
From messy signal to a useful guess.
When a photo, sound or sentence has too many details for a hand-written rule, a neural network breaks the job into small transformations and passes the signal forward.
A stack of small pattern-finders.
A neural network is a chain of simple mathematical transformations. Each layer combines incoming signals, keeps what seems useful, and hands a new signal to the next.
the basic story
input → operation → output
-
01 / input
Numbers go in.
A cat photo becomes pixel values. To the model, every image starts as a stream of numbers.
-
02 / operation
Clues get combined.
Learned weights give some signals more influence. Layers turn edges into shapes, then shapes into richer features.
-
03 / output
A ranked guess comes out.
The last layer combines the clues and returns likely answers, such as cat, fox or dog.
Pixel → edge → ear → cat.
One dark-to-light change can activate an edge detector. Several edges can form a shape; combined shapes can support a cat guess. No single unit spots the whole animal.
01 See the anatomy
A network is a stack of decisions.
Do not picture a tiny brain. Picture a relay team: each layer notices something different and hands its signal to the next.
Send a signal through the layers.
Choose what comes in, then let the network do its small, repetitive work. Click a neuron to inspect the question it is asking.
inputhidden layersoutput
hidden layer / filter 03This neuron looks for ears.
Every filter sees a small fragment. Together, the fragments make a pattern.
signal fragmentsedges / shape / context
output distribution86%
Each neuron asks one small question. Together, they spot a pattern.
02 Zoom into the filters
What does each layer notice?
Early layers catch simple shapes. Later layers combine them into things that feel meaningful to us.
Turn the filter dial.
The network does not jump from pixels to “cat” in one step. It builds a useful guess from increasingly specific clues.
activation mapline detector08 / 16 cells lit
At the start, the signal is mostly edges: light against dark, line against background.
The “knowledge” is spread across connections. No single neuron knows the whole picture.
03 Build the architecture
More layers, more steps.
A neural network is assembled as a path: one input boundary, a stack of hidden layers, then one output. Add depth and follow what each extra step can combine.
Build one and follow the signal.
Choose an example, change the number of hidden layers, then click a column. The percentages are illustrative activation strength: how loudly each layer is responding to this signal.
3 hidden layers / deep5 layers / 16 units
each column is one transformationsignal moves left → rightsignal moves top → bottom
architecture
hidden layers
run an examplecat photo
Simulation note A real model has learned weights and many more units. This compact diagram keeps only the causal story visible: what enters, what a layer combines, and what leaves.
Depth is not magic. It is more small transformations in a row.
01 / check your mental model
Which sentence feels closest?
A neural network is best described as…
Pick one. The lab will tell you if the picture clicks.
FAQ Neural networks / quick answers
One more
layer.
Three short answers to the questions people usually ask after the first experiment.
01 / definitionWhat is a neural network?
A neural network is a sequence of weighted mathematical transformations that turns input numbers into a prediction. Each layer combines signals and passes new values onward.
02 / learningHow does a neural network learn?
Training compares a prediction with an example, measures the error, and adjusts the weights so similar future predictions can improve. This lesson keeps the weights fixed so the forward path stays visible.
03 / depthWhat do hidden layers do?
Hidden layers are intermediate workspaces. Early layers can combine simple features, while deeper layers can combine those features into more specific patterns; they are not a database of human-readable rules.