What a neural network is

An expenses app reads photos of receipts. It finds the shop, the date and the total, even on a crumpled receipt in bad light.

Nobody wrote a rule for every font and every crease. There are too many.

The app uses a kind of model built from many small, simple parts. Each part does a little sum. Together they do something no one part could.

A neuron takes in a few numbers. It multiplies each one by its weight, then adds the results.

If the total is big enough, the neuron switches on. If not, it stays off.

A big weight means "this input matters a lot". A weight of 0 means "ignore this input".

# One neuron deciding: should this support ticket go to a person?
# Inputs: angry words found, times the customer has rung, order value in tens of pounds.
inputs = [3, 2, 5]
weights = [2, 3, 1]   # How much each input matters.
threshold = 15

total = 0
for i in range(len(inputs)):
    total = total + inputs[i] * weights[i]

print("Weighted total:", total)
if total >= threshold:
    print("Send to a person")
else:
    print("The chatbot can answer")

# Change the first weight from 2 to 0, then press Run.
# The neuron now ignores angry words. What does it decide?

One neuron can only make a simple choice. So we join many.

The first layer takes in the raw numbers, such as the brightness of each dot in a photo. Its answers feed the next layer, and so on. The last layer gives the answer.

Early layers spot simple things, like edges. Later layers combine them into shapes, then letters, then words.

At first the weights are random, so the answers are poor.

In training, the network sees an example and gives an answer. If it is wrong, every weight is nudged a little, so the answer would have been closer. Then the next example.

This repeats millions of times. It needs a lot of examples and powerful chips, which is why deep learning costs real money to train.