4 ms·
Where is the table comparing the accuracy of DL tools to others? Seems like mainly he is just saying "DL takes a shitload of resources" and no numbers. My und
by ilaksh 11y ago
Where is the table comparing the accuracy of DL tools to others?
Seems like mainly he is just saying "DL takes a shitload of resources" and no numbers.
My understanding is that the other tools have limited relevance now with DL, since most systems are moving to cloud-based services that do have resources for DL, and DL offers much better accuracy. That is not readily concluded from this report.
- szilard 11y agoI did not get to DL yet (did just RF+GBM so far). Stay tuned...
- rjdagost 11y agoDeep learning is definitely "the new hotness" for machine learning but it is a big mistake to assume that all other techniques are now irrelevant. Depending on your problem space, DL might produce decidedly inferior results. For example, if you do not have a massive database DL probably isn't going to give you the best results. Or, if you have limited computational resources DL is probably not the best choice.
- Fede_V 11y agoActually, resources are the least of your concerns when it comes to DL. DLs have a shitload of hyper-parameters (effectively, the entire architecture) which is mostly built by trial and error + intuition. There was some really exciting work on reversible SGD by an applied math group at Harvard (https://github.com/HIPS/hypergrad https://github.com/HIPS/hypergrad) to obtain derivatives with respect to network architecture, but it's not very mature yet. Further, DLs need an immense amount of data to be efficient. If your dataset has only a few hundred of features, and you have a few thousand samples, it's unlikely DL will help you much. Finally, DL is not really useful out of the box, at all. scikit-learn's selling point is that it has an amazingly good uniform API, fantastic documentation, and very clean code. Making a working DL implementation with a simple 'fit' sklearn-like API is impossible.
- ilaksh 11y agoI noticed you also did not give any information on the accuracy of DL versus other techniques.