Labellio "deep learning" AI image recognition test: pizza

I just finished translating an article for THE BRIDGE about some new AI tech called Labellio from a Japanese company called Alpaca that uses "deep learning" to create an AI model that can perform image recognition. What I thought was especially cool about this is the huge range of possible uses because the tech can be used to create pretty much any kind of dataset for image recognition and downloaded and implemented into applications. 

I decided to give Labellio a try and see if it can learn to differentiate different pizza toppings. So first I created a new data set called "Pizza Toppings" and put in 5 different varieties of pizza.

Labellio then automatically image searched each type in the data set which resulted in beautiful cascades of pizza that made me hungry.

Then some processing and analyzing happened which took about 5 minutes and produced this graph that I didn't really understand but I guess it has something to do with how well the types in the data set were able to be parsed? or something. From there you can test out the new data set's accuracy or download it.

To test to see if this actually works or not I google image searched each type of pizza, chose an image randomly, and copied and pasted the image link into Labellio. The results were mixed but I was pretty surprised by how accurate it was.

White, pepperoni, and supreme were all recognized by the data set perfectly. Meat lover's was recognized but it's a little unsure. The only real miss was Hawaiian, which is surprising, I thought the pink ham and bright yellow pineapple would be a dead giveaway. Overall I think this is pretty cool. Alpaca is really opening up image recognition and making it widely available to anyone.

p.s. Why did I choose pizza? Because I like pizza, and also because I thought it would be more challenging than like humans vs. dogs, but not too impossible like real leather belts vs. artificial leather belts. Sometimes I can't tell the difference. Anyway, looking forward to playing with this some more and making different data sets.