Hacker News Reader

It is interesting that the words "production ready" can be used in connection to image classifiers. Google is unarguably the leader of this in the world, and yet Google photos search for cats shows me dogs and a search for dogs shows me cats and sometimes rats. We're probably still a few years away from production-ready image classifiers of any sort.

View on HN

The definition of "production ready" for machine learning is when it's at an accuracy level that is high enough that it becomes "useful" for some sense of the word. Machine learning models will almost never achieve 100% accuracy, so it comes down to what a "useful accuracy" level is to you.

Google doesn't run those images through a pixel-level ConvNet, Google's algo uses a combo of the HTML info surrounding the image and the image's popularity. Running every image on the web through a CNN would be computationally infeasible

ConvNets for image classification (like what the article is about) often achieve better-than-average-human performance

It DOES run your photos that you store in google photos through (at least according to public statements)

> better-than-average-human performance

There is false, except under some contrived tests. NNs still aren't very good at handling rotations, can be confused by changing single pixels and often recognise shapes out of complete static.

Reply on news.ycombinator.com