6 ms·
Machine Learning with TensorFlow Enabled Mobile Proof-Of-Purchase at Coca-Cola
- amelius 9y agoJudging from the comments section, Google needs to improve their ML for scam-detection.
- jj12345 9y agoIt's a hard problem. One person's scam is another person's business opportunity. I, for one, want to learn more about this amazing spell-casting.
- sdenton4 9y agoConsider that the Latin character set doesn't change when it realizes that a new convnet is doing a good job at recognizing it... Machine learning is hard. Adversarial machine learning is much harder.
- adamzd 9y agoWhy not just use QR codes and Google Vision API? Kinda seems like reinventing the wheel here..
- shp0ngle 9y agoLate stage capitalism
- vog 9y agoWhy "late stage"? 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".
- pennaMan 9y agoIt's just a meme people like to spew when capitalism gives us some heavy irony. For example a fossil fuel company pivoting to climate change themed toys. It is irrelevant to the article at hand.
- tanilama 9y agoAs useful as Jian Yang's hot dog detector.
- speps 9y agoKind of ironic that 2 comments on that article are spam... Maybe spend some time fixing that too.
- tw1010 9y agoMan, software engineering at the cutting edge is getting harder and harder by the day. Not only are you expected to master coding but also math heavy ML and economics. I guess a consequence of software eating the world is that more and more fields of human knowledge is folded into the engineering world. The consequence being that engineers really aught to be very broad in their reading if they want to take advantage of all the low hanging fruit the octopus arms stumbles into.
- candiodari 9y agoMath heavy ML is simple: ∀ problems: solution is Convnets (CNN in the article). 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.
- ska 9y agoMath heavy ML is simple: ∀ problems: solution is Convnets Oh, if only that were true. The articles application, however, is a good example of where they are very applicable.
- samfisher83 9y agoI think software engineering is easier now than back in the day. I have heard from people who used to have to write the code by hand change it to assembly and then program their chip. Debugging was a pain. Everything was a pain. We have also these crazy tools now. We don't need to know mnemonics.
- tw1010 9y agoI agree with you. I suppose what I should have said is that software engineering is becoming much more lateral than ever before. Though it seems like the verticalness is decreasing. Hmm, I wonder if the total "area" of "verticalness" times "lateralness" has stayed constant over time :) Perhaps it is a reflection of how much time an engineer can spend on learning things, and I would expect the "area" is different across the distribution of programmers. But maybe the average area has remained invariant.
- JustFinishedBSG 9y agoCongrats google on finally inventing OCR on fixed fonts.
- AndrewDucker 9y agoOn curved surfaces, under varying lighting conditions. 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.
- nerdponx 9y agoStill good to have a healthy level of cynicism towards Deep Learning Hype articles.
- Sujan 9y agoAI at Coca-Cola... and it actually makes sense. Nice!
- fredley 9y agoThis is undoubtedly incredibly impressive as a feat of engineering. Detecting codes on those sample images with >95% accuracy is no mean feat. No doubt Google uses its muscle for many worthier things than this, and the learning from this project will be applied to many more 'serious' problems. I think what this blog post triggers for some people is just a further step down the road from Google as it was - a tech first solutions later company that solved interesting problems because they were interesting - to Google of today/the future - a big corporation that acts pretty much like any other, but with a few neat tools in its toolbox.
- amelius 9y ago> This is undoubtedly incredibly impressive as a feat of engineering. 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?
- fredley 9y agoYes, and it runs on an image with a shadow occluding half the cap, held in someone's shaky hand, with low-contrast colouring. AFAIK, postal scanning is done in good lighting with known optical parameters of blue/black on white. The main part used for sorting the postal/zip code, which is relatively clear and in caps/numerals. Almost all mail has printed labels. Accuracy rates are not published, many people are hired to sort through the ones that cannot be automatically classified. It may not seem like a huge jump conceptually but in engineering terms it's a big deal.
- vog 9y ago> No doubt Google uses its muscle for many worthier things than this 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.
- corporateslave3 9y ago
- rahimnathwani 9y agoReally nice. 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
- danjf 9y ago+ saving of millions by not having to replace old dot-matrix style printing equipment
- davidkuhta 9y agoAlso worth mentioning: - 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)
- candiodari 9y agoThis is where the current state of the art in industry is in machine learning : do "modern" things like scanning documents/codes/... on old processes without modifying the processes. 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.