Machine Learning with TensorFlow Enabled Mobile Proof-Of-Purchase at Coca-Cola
developers.googleblog.com
developers.googleblog.com
But yes, you should learn to apply multilayer convnets to image problems.
And don't worry about architecture. There's 2 variables : number of parameters and number of layers. There's an optimum, which will be bigger than "advanced" architectures, but it will match the advanced architectures' performance very closely (1-2% worse perhaps).
Slight issue is that it takes 2-3 hours to test a (number of free parameters, number of layers) combination, and so "hyperparameter tuning" (pick random and test, repeat) takes a long time. But if you make it a nice pipeline you can do it on cloud in parallel on a regular basis (as new data gets added) and have extremely good results.
Math heavy ML is simple: ∀ problems: solution is Convnets
The articles application, however, is a good example of where they are very applicable.
> We don't need to know mnemonics.
Every specialized discipline has jargon and technical terms. Today everyone memorizes Javascript frameworks instead.
> We have also these crazy tools now.
Yes, but we also have crazy tools that rapidly create complexity. The popularity of using tools we don't understand, didn't (directly) write, that we cannot meaningfully inspect/audit (e.g. modern machine learning) is rapidly adding unknown, interdependent complexity. This complexity is already[2] spinning out of control. The only reason it seems easier today due to most of the problem being ignored.
It probably depends on your industry though, working on new experimental projects has a lot of benefits now because the experiments are talked about in the open on forums like Reddit and Github
I hope this doesn't sound too negative, but since Google's main revenue is ads, "worthier things" ultimately means "improved ad targeting", doesn't it? In that case, it is not so different from the demonstrated application.
For years, post offices around the world have been automatically recognizing handwritten text, which is of much poorer quality and consistency than those codes.
Not sure what is new here. That it runs on a smartphone?
It may not seem like a huge jump conceptually but in engineering terms it's a big deal.
Machine learning is hard. Adversarial machine learning is much harder.
Say you want to do the famous quality check in factories (is the dogfood box closed properly ?), you can just do that with a convnet.
tl;dr:
- Customised text recognition of codes printed on goods
- UX flow designed to (i) help users correct errors, and (ii) gather additional labelled images
- Fast: Needed a one-second average processing time (product-code iamge -> OCR pipeline)
- Accurate: Goal was to achieve 95% string recognition accuracy (with improvement via active learning)
Nobody knows how long the current capitalism will stay alive. This might as well be "early stage" ... in the sense of "one of the many experiments of companies to find the best form of market segmentation".
If you want to show that standard OCR can cope with the images they have on that page, then go for it. But I suspect you'll have issues.