Perceptron
The first trainable neural network, from 1958 — and the story of how a book killed it is the most repeated wrong story in AI.
When not to use it
- (Nobody uses a perceptron. The question is what the story teaches.)*
- As evidence that critics kill fields. They identified a real unsolved problem and it took seventeen years to solve.
- On non-separable data. It cycles forever. No graceful degradation.
- As a model of a modern neuron without the caveat. The hard threshold is exactly what blocked gradients.
- Citing "the XOR thing" without reading them. The order/diameter results are the real content.
Reach for something else instead
- Logistic regression — a perceptron with a sigmoid and a probabilistic interpretation. Strictly better.
- SVM — maximises the margin the convergence theorem depends on.
- Multilayer networks — what Minsky and Papert said would work if you could train it. You can now.
Sources & further reading
- Rosenblatt (1958), The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain — the machine that learned.
- Minsky & Papert (1969), Perceptrons: An Introduction to Computational Geometry — read what they actually claimed; the order/diameter results are the substance.
- Rumelhart, Hinton & Williams (1986), Learning representations by back-propagating errors — the answer to the real objection, seventeen years later.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Repeating the myth that Perceptrons killed neural networks out of dogma. They named a real problem: nobody could train multilayer nets.
- Thinking XOR was the book's main result. The scaling limitations are the mathematics.
- Missing that the hard threshold was the actual obstacle. Swapping it for a sigmoid is what unlocked backprop.
- Reading the 1958 press coverage as a quaint historical oddity. The overclaiming pattern is unchanged.
At a glance
FieldDeep Learning
Year1958
What it wasthe first machine that learned its own weights from examples
The famous storywrong; Minsky & Papert knew multilayer nets could do XOR
Their real claimnobody could train them, which was true until 1986
The actual obstaclethe hard threshold has no gradient
DifficultyBeginner
Flashcards for this concept
Question
—
Answer
—
1 / 4
Often compared with
Perceptron vs. modern neuron — the same weighted sum and threshold, with a smooth activation instead of a hard step. That one substitution is the difference between 1958 and everything.