AIToday
AI Safety & AlignmentZenn AI/MLPublished: Nov 18, 2026, 04:00 JST

XOR defeats perceptron: 2/4 in 1,000 epochs, as in 1969

Send on LINE
XOR defeats perceptron: 2/4 in 1,000 epochs, as in 1969

3 Key Points

  1. What happened

    Running the same perceptron code on AND and XOR, AND scored 4/4 but XOR scored only 2/4 after 1,000 epochs. A single straight line can't separate XOR's 1s and 0s, a limit Minsky and Papert set out in 'Perceptrons' in 1969.

  2. Why it matters

    That 2/4 result shows the first AI winter's trigger was a structural ceiling, not a funding shortfall. When a single line is the only shape the model can draw, no amount of training fixes the problem.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Context & Analysis

The XOR result sits inside a longer pattern. Before the first winter, the boom centered on inference and search, but search could only solve toy problems, and the simple perceptron could not handle XOR; the 1969 book 'Perceptrons' and the UK's Lighthill report in 1973 are cited as the triggering events. The second winter, in the late 1980s through the 1990s, had a different center: expert systems that encoded expert knowledge as IF–THEN rules. That approach stalled on the knowledge-acquisition bottleneck, since experts cannot always articulate their tacit judgment, and rules need constant revision as conditions change. The collapse of the dedicated LISP machine market and the end of Japan's Fifth Generation Computer Project in 1992 mark that downturn. Both winters share one shape: the limit was a form decided in advance by people — one straight line in the first case, hand-written rules in the second. The third boom's machine learning and deep learning are framed as a move toward learning that form itself from data, which ties the two historical winters to the technical shift that followed.

FAQ
Why can't a perceptron learn XOR?
A perceptron can only draw one straight line on the input plane. XOR's 1s and 0s sit diagonally, so no single line separates them.
How do you make XOR learnable?
A hidden layer builds XOR from (x1 OR x2) AND NOT (x1 AND x2). The open problem was how to learn those middle-layer weights from data.
What ended the first AI winter?
Backpropagation became widely known in 1986 through a paper by Rumelhart, Hinton and Williams, answering how to train the middle layer.
Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next articleAIエージェント guide: delegate email and research tasks, not whole jobs