The Eastbourne projectTeaching one neuron to spot frosty months
Load the station record, train a perceptron on the old years and test it on the new ones.
The learner downloads the Eastbourne station file and reads 812 months, January 1959 to August 2026. Three markers need care before any learning starts. The latest 8 months are labelled Provisional and are left out. 88 months carry an asterisk meaning a value was estimated, which the learner keeps but records. And a hash sign on sunshine marks a change of instrument. Each month is then labelled frosty if the station logged at least one day of air frost. The months up to 2004 become training data, 552 of them; 2005 to 2025 are held back as a test, 252 months. Splitting by time matters: the neuron must never learn from the years it is later tested on.
Frost-month predictions for Eastbourne, our Python run on Met Office station data, 28 September 2026| Method | Training accuracy (1959 to 2004) | Test accuracy (2005 to 2025) |
|---|
| Always say "no frost" | 68.7% | 75.4% |
| Perceptron, minimum temperature and sunshine, raw numbers | 81.9% | 79.4% |
| Perceptron, same two inputs, scaled first | 91.8% | 89.7% |
| One rule: frost if minimum temperature is 5.2°C or lower | 90.0% | 91.3% |
The first comparison is with doing nothing clever: a program that always answers "no frost" is right 75.4% of the time on the test years, simply because most months are frost-free, so any model has to beat that. The second lesson is scaling. Sunshine is measured in hundreds of hours and temperature in single degrees, so with raw numbers the sunshine input swamps the weight updates and the neuron manages only 79.4% on the test years. Rescaling both inputs to a common size before training lifts that to 89.7%. Nothing about the neuron changed except the units of what it was fed.
The third lesson is the most important. The perceptron keeps making mistakes on every pass through the data, 62 errors on the last of 100 passes, because no straight line can separate frosty from mild months perfectly: some cold months have no frost and some mild ones do. A single neuron can only draw straight lines, which is why real AI systems stack many of them. And the plainest model of all, one rule on minimum temperature, scores 91.3% on the test years, slightly better than the neuron. Knowing when a simple rule is enough is a skill every AI builder needs.
Ages 8 to 11
Sort weather cards into frosty and not, then find one rule that gets most of them right.
Ages 11 to 15
Read the Met Office file in Python and test simple rules against the baseline.
Ages 15 and up
Code the perceptron, scale the inputs and compare it fairly on years it never saw.
Met Office data, our neuron
Monthly observations come from the Met Office historic station data for Eastbourne. The labels, the training, the test split and every accuracy figure on this page are our own work.
