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
In the real world, you don't need the extra 2% accuracy on CIFAR10. You just need a pretty good model. And you can train that for a few dollars on a cloud platform. Computing is not a bottleneck today. https://twitter.com/Java07/status/969344010589437953 …
The fact that many papers use absurd amounts of compute is an artifact of the publishing model combined with the availability of these resources at large companies. In the real world, you only need massive compute if you have massive data -- and that implies you can afford it.
Maybe you generate massive data, e.g. when training the model to do things. If I wanted to have a pro-level neural network playing Go? Compute is definitely the bottleneck.
Yesterday I had to throw out 99% of my data because training on the full set required a few hundred gigabytes for a single numpy array. So not always true, but a good point nonetheless!
dask.distributed alternately, many implementations of hdf5 can dynamically load necessary data from disk also online models in general -- fitting a batch model on hundreds of gigabytes of observations at once is not typically how these things are done
Don't forget time spent developing a better model. Devs are far more expensive than servers. It is important to be able to determine how much refinement is enough.
Indeed, I see myself facing the challenge of too little Data much more often than that of too little computational power.
Sometimes you need to do absurd amount of compute to find that one simple model which has the accuracy that you need but can still fit into the hardware constraints of the platform you are deploying into. This is a big issue in deploying ML models on tiny edge devices!
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.
