One neuron: w·x + b and an activation
Every neural network is built from one simple unit: multiply the inputs by weights, add a bias, pass the sum through an activation. Here is that unit on its own, with its two inputs drawn as a plane.
Interactive simulation. It needs JavaScript; the text below explains what it shows.
What you are looking at
The colour shows the neuron's output for every point (x₁, x₂). The white line is where w·x + b = 0, the neuron's decision line; the amber arrow is the weight vector, which always points across the line toward the high side. Drag the white probe to see the calculation for one point.
Things to try
- Change only b: the line slides without turning.
- Change w₁ and w₂: the line turns; their size sets how sharp the transition is.
- Switch between step, sigmoid and ReLU: same line, very different outputs.
- Make both weights zero: the neuron ignores its inputs and outputs a constant.
How it works in real systems
A single neuron is logistic regression when the activation is a sigmoid, and a perceptron when it is a step. Real networks stack thousands of them so that many straight lines combine into any shape.
Code lab
The same idea in real code, step by step. Copy each block into a notebook or a file and run it in order.
Step 1 · A neuron is three lines
import numpy as np
w = np.array([1.2, -0.8]); b = 0.2
def neuron(x, act="sigmoid"):
z = w @ x + b # weighted sum plus bias
return {"step": float(z > 0), "sigmoid": 1 / (1 + np.exp(-z)), "relu": max(0.0, z)}[act]
print(neuron(np.array([0.4, 0.3]))) # z = 0.44 → sigmoid 0.608, as in the simulation
Step 2 · The decision line
The neuron's output is 0.5 exactly where z = 0, the line w₁x₁ + w₂x₂ + b = 0. The weight vector is perpendicular to it; the bias shifts it.
import matplotlib.pyplot as plt
xx, yy = np.meshgrid(np.linspace(-1, 1, 200), np.linspace(-1, 1, 200))
Z = 1 / (1 + np.exp(-(w[0] * xx + w[1] * yy + b)))
plt.contourf(xx, yy, Z, 20, cmap="Blues"); plt.contour(xx, yy, Z, [0.5], colors="k"); plt.show()
Step 3 · Train it: logistic regression
One sigmoid neuron trained with cross-entropy is logistic regression.
from sklearn.datasets import make_classification
X, y = make_classification(200, n_features=2, n_redundant=0, random_state=1)
w, b, lr = np.zeros(2), 0.0, 0.1
for _ in range(1000):
p = 1 / (1 + np.exp(-(X @ w + b)))
w -= lr * X.T @ (p - y) / len(y); b -= lr * (p - y).mean()
print(w, b, ((p > 0.5) == y).mean())