4 ms·
There’s no such thing as a free lunch. In the case of deep learning - we’ve figured out as an industry how to apply a particular model of linear algebra to solv
by mikehollinger 8y ago
There’s no such thing as a free lunch. In the case of deep learning - we’ve figured out as an industry how to apply a particular model of linear algebra to solve really interesting problems. It does way better than you’d guess in particular spaces. Hotdog or not hotdog is a perfect example of this. I actually (for fun) pulled several thousand pictures of hotdogs from the MS COCO dataset and trained a model to do that. It did well - but what I’d actually created wasn’t a hotdog detector - is created a bun detector. I could prove this by giving it an image of a sandwich and the model would confidently score it as a hotdog.
We aren’t mimicking how our brains work in any broad sense. At best it’s a very narrow definition.
In this case - the larger the matrices - the more you can do. There’ll probably be a Moore’s law of AI at some point.
- wetpaws 8y agoGiven how much BS people typically believe, I would say a problem of training data is universal and not something exclusive to neural networks :)
- Baeocystin 8y agoSounds like a real example the apocryphal tank/shadow detector. https://www.gwern.net/Tanks https://www.gwern.net/Tanks
- gwern 8y agoIf you showed a human who'd never heard of hotdogs before thousands of photos of hotdogs, and every single one of them had buns, why do you think they would assume that a naked wiener would be the true definition of hotdog, rather than 'bun with meat/filling'?
- Baeocystin 8y ago>and every single one of them had buns Why would you assume that? I did a cursory GIS and Bing search, and somewhere between 1-5% of the returned images were bunless wieners.
- gwern 8y agoHe didn't use GIS/Bing, or anything derived from them like WebVision. He used MS COCO. And if it had really returned 5% naked wieners, the results he described would be a lot less likely.
- Baeocystin 8y agohttp://cocodataset.org/#explore http://cocodataset.org/#explore Literally the first image returned for hot dogs is a bunch of bunless ones on a grill. There are plenty more.
- gwern 8y agoThen maybe he didn't train it as well as he thinks he did. For comparison, I tried out Clarifai's image classification API: https://clarifai.com/demo https://clarifai.com/demo It has no problem classifying some bunless ones on a grill (https://www.hot-dog.org/sites/default/files/2016-09/sausage%20on%20grill.jpg https://www.hot-dog.org/sites/default/files/2016-09/sausage%...) as being sausage/pork/beef/hotdog, with nary a 'sandwich' to be seen, while conversely, a Big Mac (https://d1nqx6es26drid.cloudfront.net/app/uploads/2015/04/04043402/product-big-mac.png https://d1nqx6es26drid.cloudfront.net/app/uploads/2015/04/04...) gets sesame/burger/lettuce/sandwich/bun/bread tags with no sausage/bratwurst/hotdog to be seen. Clarifai's NNs seem to do fine on 'hot dog or not hot dog'...