Perceptron
The first learning artificial neuron: a weighted sum of inputs followed by a hard yes/no threshold, with a simple rule to learn the weights.
Rosenblatt, 1958
A tiny judge that listens to a few clues, gives each clue some importance, adds it all up and says yes if the total is big enough. When it gets it wrong, it adjusts the importance a little.
Where it fits
+All relationships as a list
Is a kind of / part of
Siblings
Evolved into
How it works
Frank Rosenblatt’s perceptron (1958) is a single artificial neuron:
- Multiply each input by a weight and add a bias.
- If the total is above zero, output 1; otherwise 0.
Learning is just as simple: when the perceptron makes a mistake, shift the weights a little toward the correct answer. If the two classes can be separated by a straight line, this rule is guaranteed to find one.
In 1969 Minsky and Papert highlighted its limits: a single perceptron cannot learn even XOR, because XOR needs a curved or two-part boundary. Interest in neural networks cooled. The fix, stacking neurons into layers and training them with backpropagation, became the multilayer perceptron and, eventually, all of deep learning.
Everyday use cases
- Case 01
A historical milestone
The perceptron showed in 1958 that a machine could learn from examples, kicking off decades of neural network research.
Strengths, weaknesses and when to use it
Strengths
- Very simple learning rule.
- Guaranteed to converge on linearly separable data.
Weaknesses
- Can only draw a straight decision boundary. It cannot learn XOR.
When to use itToday mainly to learn how neural networks started. Its multi-layer descendants are what people use.
Compare with
Check yourself
- Q01
Which simple function can a single perceptron NOT learn?
Show answer
C: XOR XOR is not linearly separable, so no single straight line splits its classes.