Neural Networks: How They Recognize Handwritten Digits

How a Neural Network Recognizes Handwriting

A gentle walkthrough — from pixels to digits, without complexity.

Why This Matters

A 28×28 pixel image of the number 3 looks like a messy grid to a computer. Yet your brain instantly sees the digit — even though the exact pixel values vary. Training a program to do the same is surprisingly hard. Neural networks offer a solution, and this summary breaks down how they work.

The Network Architecture

From Pixels to Answer

Input layer: 784 neurons (one for each pixel in 28×28). Activation = grayscale value (0 = black, 1 = white).
Hidden layers: Two layers, each with 16 neurons. (Choice is illustrative; many variations exist.)
Output layer: 10 neurons, one per digit 0–9. The brightest neuron is the network’s guess.

How Information Flows

Activations in one layer determine activations in the next.
Each connection from a neuron in layer n to a neuron in layer n+1 has a weight (a number).
Each neuron in the next layer also has a bias (another number).
The weighted sum of inputs + bias is passed through an activation function (e.g., sigmoid or ReLU) to produce the neuron’s activation (0–1).

Key Numbers in the Network

784 input pixels

2 hidden layers

16 neurons per hidden layer

10 output digits

~13,000 weights & biases total

Training a network means finding useful values for all ~13,000 parameters so that the network correctly classifies handwritten digits.

Weights, Biases & Activation

Weighted sum: Multiply each input activation by its connection weight, then sum everything. Add the bias.

Sigmoid (logistic function) squashes any real number into [0,1]. Used in early networks; mimics “on/off” neurons.

ReLU (Rectified Linear Unit): output = max(0, input). Today more common — easier to train in deep networks.

“ReLU is inspired by a biological threshold: a neuron fires only if the input is strong enough; otherwise it stays at zero.” — Lisha, ML researcher

What Do Hidden Neurons “Learn”?

Ideally, the network builds up compositional features:

Layer 2 detects small edges (e.g., oriented edges in specific positions).
Layer 3 combines edges into larger patterns (loops, long lines, corners).
Output layer maps pattern combinations to digit labels (e.g., a loop on top + a line on the right → “9”).

This hierarchical decomposition is plausible and gives the network a logical structure — not just a black box.

Insights to Remember

“Each neuron is just a number between 0 and 1. The whole network is one big function with ~13,000 tunable knobs.”

“If you had to set all weights and biases by hand, you’d go mad. Training automates that — finding useful parameters from examples.”

Understanding these internal parameters helps debug, improve, and even challenge our assumptions about what the network is doing.

What’s Next?

Training is the subject of the next video: how the network learns the right weights and biases by processing examples. Expect deeper dive into backpropagation and gradient descent.

Meanwhile, explore linear algebra (matrix multiplication) — it’s the engine powering all these calculations.

Takeaway

Neural networks are inspired by biological neurons but mathematically simpler. A basic feedforward network with a few hidden layers can already recognize handwritten digits with surprising accuracy. The key lies in the weights, biases, and activation functions — and the ability to train them from data. This is not true intelligence, but a powerful pattern recognizer built on layered abstractions.