When we embarked on this journey, we thought, “this seems plausible, why wouldn’t we do it?”
It was when we started looking at building the thing, and you realise what a complex beast this really is.
Unlike Vivino, which uses machine learning to read text from labels, cheese has not always got a label. In fact, when you mostly snap a shot of some cheese, it is in it’s element.
Cheese also has various states in which it appears depending on it’s maturity. So what does that mean for the technology?
In order to try and get an accurate model of all cheeses in existance (over 4000, apparently) you would then need roughly 10,000 images of each cheese to ensure the model really knew the small differences between each cheese.
We trained Cheezus on 10 initial cheeses which took over 9000 images. We now have over 150 cheeses on the database, but it’s still only learning all of these.
It will take time for the machine learning to learn every cheese in the world, but that’s where we need our cheese lovers to help. Take as many pictures as you can of a cheese, any cheese; if Cheezus doesn’t recognize it, simply correct it and submit.
Every time you do, Cheezus learns.