Inside a Neural Network: How Machines Learn Patterns
Peek inside the brain of a machine
Peek inside the brain of a machine
A neuron takes inputs, multiplies each by a weight, adds them up, and passes the result through an activation function. That is all it does. Drag the weight sliders to see how a single neuron processes two inputs.
One neuron cannot do much. But connect them in layers and they can recognize anything. The first layer detects simple patterns. Each subsequent layer combines those into more complex features. Click on any neuron to see its activation.
Click any neuron to highlight its connections.
Raw data goes in: pixel values, audio samples, text tokens. No processing here.
Where the magic happens. Each layer extracts higher-level features from the previous layer.
The final answer: "cat" or "dog", spam or not spam, a number from 0 to 9.
A neural network starts with random weights and makes terrible predictions. Training shows it examples, measures how wrong it is (loss), and adjusts weights to be less wrong next time. Click "Train" to watch the loss decrease.
GPT-4 was trained on hundreds of billions of words. Training ran on thousands of GPUs for months. The electricity bill alone was estimated in the millions of dollars.
The neurons and layers you just explored power the AI behind everyday technology.
Your phone recognizes faces, identifies plants, and reads text in photos using convolutional neural networks (CNNs).
ChatGPT, translation apps, and autocomplete all use transformer networks trained on text to predict the next word.
Multiple neural networks process camera, lidar, and radar data simultaneously to detect lanes, pedestrians, and obstacles.
Neural networks can detect cancer in X-rays, predict protein structures, and identify drug interactions.
You've seen how simple math, repeated millions of times across connected layers, lets machines recognize images, understand speech, and even generate text. Every AI application is built on these ideas.
Each neuron just multiplies inputs by weights, adds them up, and outputs a number. That is it.
One layer detects edges. The next detects shapes. The next detects objects. Simple units create complex understanding.
The network starts with random guesses. Training nudges weights so the output gets closer to the right answer each time.
A neural network is only as good as its training data. Biased data creates biased AI.
Put your new knowledge into practice!