The bus lesson that makes a learning machine less jumpy
The night bus rolled onto a dark bit of road where the lane paint had vanished. The driver tightened his grip. In training, the instructor sometimes sticks removable patches on the windscreen, blocking random bits of the view, so the driver learns to steer using lots of clues at once.
A driver trained only on perfect roads can get cocky and follow one easy hint, like a single sign. When that hint changes, the bus drifts. A big pattern-finder used in things like reading handwriting can do the same, clinging to tiny quirks from practice instead of the real shape of things.
The new training habit is bold but simple. While the pattern-finder is learning, some of its little inside workers are randomly switched off for each example, so they stay quiet for that moment. On the bus, it matches the instructor changing which windscreen patches block the view every run.
That random blocking breaks lazy teamwork. A worker that only does well when its favourite partner is present can’t rely on that partner any more. The driver has to steer whether the left edge shows up, the right edge shows up, or only the far horizon is clear. Takeaway: practise with missing pieces so you don’t panic when one clue disappears.
There’s a catch when it’s time to decide for real. You want one steady answer, not a jumble of half-switched parts. So the full set is used, but each connection is turned down to match what the system got used to during patchy practice. It’s like driving with a clear windscreen, yet not trusting any single cue too much.
Later, on that unmarked stretch, the driver didn’t hunt for one special landmark. He held the bus centred by weighing many small hints together. That’s the point of the training habit: not a new bus or a new road, just practice that stays calm in messy conditions. Tools that sort photos or tidy up speech can end up steadier for it.