One neuron, one weighted decision
A neuron combines inputs with learned weights, adds a bias, and applies an activation function.
The idea
Several clues each get a different importance. Add their weighted effects, include a baseline, then pass the result through a response rule.
For a neuron, z = w·x + b and a = φ(z). A multilayer network composes many such transformations, and backpropagation supplies gradients used to update parameters.
How it works
- Compute weighted sums plus bias.
- Apply an activation function.
- Propagate activations forward to produce a prediction and loss.
- Backpropagate gradients and update parameters with an optimizer.
One dense neuron
z is the affine pre-activation; φ is a nonlinear activation such as ReLU, tanh, or sigmoid depending on architecture and purpose.
z = w·x + b, a = φ(z)
Check your understanding
What is learned during neural-network training?
- Weights and biases
- Only the page theme
- The browser width
- Only the test labels
Answer and explanation
Training updates parameters such as weights and biases using gradients from the loss.
A common misconception
A neuron is not independently interpretable as one human concept. Hidden units are learned numerical features whose meaning can be distributed across the network.
Deep Learning
Build intuition for weighted neurons, shared filters, recurrent state, and self-attention before opening the full labs.