The Pontypool projectTeach it the north, then the south, and watch the north disappear
Postcode positions, five wards, a small neural network, and a test of what it still remembers.
Every NP4 postcode in Torfaen comes from Ordnance Survey's Code-Point Open, with its map position and its ward. Keeping only those inside the Pontypool built-up area leaves 721 postcodes in five wards. The model's job is simple to state: given a postcode's position, say which ward it is in. A quarter of each ward is held back for testing, and the network, two layers of 64 neurons written with scikit-learn, never sees those during training.
The lesson comes in two parts. Task A is the three northern wards, Pontnewynydd and Snatchwood, Pontypool Fawr, and Trevethin and Penygarn. Task B is the two wards whose postcodes lie furthest south on average, New Inn and Panteg. The network learns task A first and is tested. Then it carries on training, but only on task B, as a real model might be updated with only the newest data. The learner runs the whole thing with 20 different random starts.
Share of held-back postcodes placed in the right ward, average of 20 runs, our Python run| Training | Northern wards (A) | Southern wards (B) |
|---|
| A only | 97.8% | not yet taught |
| A, then B only | 0.1% | 99.3% |
| A, then B with 30 A points replayed | 92.0% | 99.4% |
| A and B together from the start | 97.6% | 99.2% |
After the second round of training the network was almost perfect on the southern wards and had forgotten the northern ones completely: on average it got 0.1% of them right, and it labelled 99.9% of the northern test postcodes as New Inn or Panteg. Nothing about the northern postcodes changed. The network simply rebuilt its weights for the new task, because nothing in its training told it the old one still mattered.
The fix is replay, sometimes called rehearsal. The learner keeps a tiny memory of the first task, just 10 postcodes from each northern ward, 30 in all, and mixes them into every round of the second training. Northern accuracy then stays at 92.0% on average, though it varied from 81.7% to 98.3% between runs, depending on which 30 postcodes were remembered. Training on both tasks at once does a little better again, at 97.6%, but only if all the old data is still to hand, which in real systems it often is not.
Ages 8 to 11
Learn five new words a day without revising old ones, then test yourself on day one's words.
Ages 11 to 15
Plot the postcodes by ward in Python and train a simple classifier on the northern wards.
Ages 15 and up
Reproduce the forgetting, add a replay memory, and measure how its size changes what survives.
Sources and limits
Postcodes and wards are from Ordnance Survey Code-Point Open 2026.3.0 (Royal Mail and OS data, Open Government Licence) with postcodes.io lookups; census figures are from the ONS. The model, the 20 random starts and every percentage in the table come from our own run with scikit-learn. A ward classifier is a teaching task, not a tool anyone needs: the official ward of a postcode is already published.
