Inside a Neural Network
Loading simulation…
flagWhat you'll discover
- arrow_forwardDescribe the structure of a feed-forward neural network
- arrow_forwardExplain the role of input, hidden and output layers
- arrow_forwardShow how information flows forward through weighted connections
- arrow_forwardConnect network depth to the complexity of patterns it can learn
Layers of neurons
A single neuron can only make simple decisions, but wire many together in layers and something remarkable emerges. A typical feed-forward network has an input layer (one neuron per feature — for a 28×28 image, 784 inputs), one or more hidden layers in the middle, and an output layer (one neuron per possible answer — say, 10 neurons for the digits 0–9).
Each neuron in one layer connects to every neuron in the next, and each connection carries its own weight. The hidden layers are where the real work happens: early hidden neurons detect simple features, and deeper layers combine those into ever more abstract concepts — edges become shapes, shapes become objects. The whole network is just millions of the tiny neurons you met earlier, wired in sequence.
The forward pass
When you feed the network an input, a wave of activity travels forward from layer to layer — the forward pass. Each neuron receives weighted signals from all the neurons behind it, sums them, adds its bias, applies its activation, and passes its own output onward. By the time the signal reaches the final layer, the output neurons hold the network's answer — a probability for each possible class.
In the 3D model you can watch this happen as light flowing along the connections, with neurons glowing brighter as they fire more strongly. The thickness or colour of each wire reflects its weight: strong connections dominate the signal, weak ones barely matter. After training, this tangled web of weights is, literally, the network's learned knowledge encoded as numbers.
Why depth matters
A network with a single hidden layer can, in principle, approximate almost any function — but it might need an astronomical number of neurons to do so. Depth turns out to be far more efficient: stacking many modest layers lets each one build on the last, composing simple features into complex ones with relatively few total neurons.
This is why the field is called deep learning — "deep" simply means many layers. A modern image network might have 50 to 200 layers; a large language model like GPT-4 has around 100. Each added layer gives the network the capacity to represent one more level of abstraction, which is why depth, more than width or raw neuron count, has driven the leap from digit-recognition to systems that can hold a conversation.