You can add location information to your Tweets, such as your city or precise location, from the web and via third-party applications. You always have the option to delete your Tweet location history. Learn more
One way to think of a neural network is as a hashtable where the hashing function is locality-sensitive. It memorizes training inputs & targets, and is capable of successfully querying targets for test inputs that are very close to what it has already seen.
In this light, the "intelligence" of the network comes purely from its training data. The network is sample-inefficient and only performs local generalization. The next frontier is abstraction & reasoning, which will enable extreme generalization and decent sample efficiency.
O boy! I agree but what a way to deflate hype. I remember pushing fast generalized classifications using hash tables in my PhD... That was 30 years ago...
Yup, all the ideas are old. The clothing is new :)
Is there any research/litterature on this interpretation?
@fchollet nice interpretation.Thanks
Yes!!!! So glad to see this phrasing. There was a ams article talking about the math of NN’s that had like one sentence of this and I spent like a week dreaming about the interpretation.
Think of it this way too. A dictionary w. a similarity hash. This is easiest to see with any NN after an argmax on softmax layer (ie sends to a bucket). For scalar valued outputs,it's function approximation. More points give more coverage for better interpolation/bucket selection
Autoencoders and GANs don't quite fit in this story though. Autoencoders, lossy compressors. GANs closer tho. GANs apply transformations, but better thought of in terms of ODEs they define and buckets become basins.
Which in a sense why they can “abstract” in a sense, but can’t learn. The learning is in the original paring of inputs and targets.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.
