Stop 07 · Deep Learning FoundationsTopics · 1/11

Neural Network Basics

In short

Phone face login, voice typing and chatbots all run on neural networks. Underneath the hype, each one is just layers of tiny calculators passing numbers forward.

Layers of interconnected nodes that transform inputs through weighted sums and non-linear functions to approximate complex patterns — the base unit of everything from image classifiers to LLMs.

The mascot sits at a desk receiving coloured rating cards passed forward by rows of small helpers.

Think of it like this

The cricket selection committee

Scouts each rate one skill: batting, fielding, fitness. Their ratings go to senior selectors, who weigh some scouts more than others. The chief selector combines those views into one decision. Each person is a neuron, and how much you trust each person is a weight.

Watch it happen

One forward pass

Inputs

Weighted sum

Activation

Next layer

Output

1/5Features come in as numbers, like pixel values or a customer's age and spend.

The big ideas

What you need to know

A neuron is tiny

Weighted sum plus bias, then an activation. That's all one neuron does.

Layers learn features

Early layers learn simple patterns, like edges. Later layers combine them into ideas, like 'face'.

Weights are learned

Training adjusts millions of weights so the output gets closer to the right answer.

Hungry for data

Networks shine on images, audio and text with lots of examples. On small tables, tree models often win.

See it in code

Walk through the code

python

1import numpy as np2 3x = np.array([0.8, 0.2, 0.5])          # 3 input features4W = np.array([[0.4, -0.6, 0.1],        # 2 neurons, 3 weights each5              [0.3,  0.9, -0.5]])6b = np.array([0.1, -0.2])7 8z = W @ x + b                           # weighted sum per neuron9h = np.maximum(0, z)                    # ReLU activation10print(z, h)

1/4One example with three features.

One layer by hand. z comes out as [0.35, -0.03], and ReLU turns the negative one into 0.

Save yourself the pain

Common mistakes

Oops

Using a neural network for a 500-row spreadsheet problem by default.

Do this instead

Try gradient boosting or logistic regression first. Neural networks need lots of data to beat them.

Oops

Feeding raw features with wildly different scales.

Do this instead

Scale inputs, for example pixels divided by 255, so training is stable.

Your mission

Play with a network in your browser

~20 min

Quick check

What does a single neuron compute before its activation?

Go deeper

1Draw a small neural network by hand and manually compute a forward pass.
2Learn the roles of weights, biases and layers before touching a framework.
3Build intuition for why depth and non-linearity let networks model complex functions.