But what is a neural network? | Deep learning chapter 1

TL;DR

A neural network is a layered mathematical function that uses 13,000 tunable numbers (weights and biases) to transform raw pixel data into a recognized output, like a handwritten digit.

Key Points

• 1.Neurons are just numbers: Each neuron holds a value between 0 and 1 representing how activated it is — 784 input neurons for pixels, 10 output neurons for digits
• 2.Layers detect increasing complexity: Early layers detect edges, middle layers detect shapes/patterns, final layers combine those into a recognized digit
• 3.Weights determine what each neuron reacts to: Each connection between neurons has a weight; the weighted sum of inputs determines if the next neuron activates
• 4.Biases set the activation threshold: A bias number shifts when a neuron fires, letting you control how strong a signal must be before it matters
• 5.Sigmoid (or ReLU) squishes the output: The weighted sum is fed through a function to keep outputs between 0 and 1; modern networks use ReLU over sigmoid for easier training
• 6.Learning = finding the right 13,000 numbers: Training the network means adjusting all weights and biases until the outputs are consistently correct — covered in the next video