Everyone says more data makes AI better. We learned the opposite the hard way: a small slice of the wrong data almost broke a model that was working.
How a collar knows a dog is eating
Our collar can't see the bowl. What it can feel is chewing. When a dog eats, its jaw sets up a small, fast, rhythmic vibration that travels to the neck, and our model looks for that rhythm: roughly two to three and a half chews a second, together with the head going down.
The model sorts every moment of the day into four things: sleep, walk, play and eat. On our test recordings it was doing well.
But it only knew chewing from a limited number of examples. Real dog recordings with every meal carefully labelled are rare. We wanted more chewing, so we went looking.
Where all the chewing data lives
It turned out the world is full of chewing data. It just isn't from dogs. Livestock researchers put motion sensors on cattle all the time, because cows spend hours every day chewing, and they publish what they record. We found an open dataset of 30 calves with their eating and drinking carefully labelled, and added a small slice of it to our training data.
The model got worse. Not just at eating: its overall score dropped.
Same word, different movement
A cow and a dog both "chew", but their jaws do completely different things. A cow grinds: a slow, steady, side-to-side motion, kept up for a long time. A dog snaps and gulps: a few fast chews, a lot of head movement, and the bowl is empty.
We had told the model that the cow's slow grinding was also "eating". So it started looking for the wrong rhythm, and began to confuse ordinary dog movements with meals and meals with other things. A few recordings of the wrong animal bent the whole picture.
We took the calves out, and the model came back. The calf dataset is still switched off in our code, with a one-line note: "Hurts dog F1 — calf signals too different".
Why the right data is gold
This changed how we think about the whole product. More data is not the goal. The right data is. One small batch from the wrong source can undo months of work, and you only notice if you measure.
That is why our app now asks owners a simple question when the collar sees something that might be a meal: "Did she eat around 18:05?" Every answer is a real dog, wearing our real collar, eating its real dinner. It is slower than downloading a dataset, and much more valuable.
- A label is not a movement. "Eating" for a cow and "eating" for a dog are different signals.
- A small amount of wrong data can hurt more than a large amount of missing data.
- Measure after every change. We only caught this because the score dropped.
We will publish how well the collar recognises eating once we have measured it on our own collar. Not before.