Lecture 1.3 — Some simple models of neurons [Neural Networks for Machine Learning]

3-Second Blueprint (English)

1. Linear neurons output a bias plus weighted inputs, whereas McCulloch-Pitts binary units fire spikes only once that sum exceeds a threshold.
2. Rectified Linear Units (ReLU) clamp negative weighted sums to zero while preserving linear scaling above zero for non-linear decision making.
3. Logistic sigmoid units map inputs onto smooth 0-to-1 curves that stochastic models interpret as firing probabilities or Poisson spike rates.