3 ms·
So to me, deep learning means learning very complex functions from data. Sometimes those functions work as classifiers, but other times they are just peculiar
by etaioinshrdlu 7y ago
So to me, deep learning means learning very complex functions from data.
Sometimes those functions work as classifiers, but other times they are just peculiar functions between two domains (imagine style transfer models or GANs).
Is Bayesian Deep Learning worth looking into if I don't have much interest in statistical applications but much more just using deep learning as general purpose functions? For that matter, what about causal inference?
- nestorD 7y ago> Is Bayesian Deep Learning worth looking into if I don't have much interest in statistical applications but much more just using deep learning as general purpose functions? If you manage to build a model with a high accuracy and a sensible uncertainty on the output, you can use that information to do a lot of great things such as : - obviously apply the method to domains that require uncertainty for legal or technical reasons (simulation) - adding sample to improve you knowledge around uncertain input (active learning) - use an optimized betting strategy that takes risks into account (bayesian optimization) Gaussian process are a prime example of that but I am not aware of a deep learning approach that realises this (yet).
- nestorD 7y agoI stand corrected, after some research, I found a paper that offers a promising way to get good uncertainty information from a deep learning model : http://www.cs.ox.ac.uk/people/yarin.gal/website/blog_3d801aa532c1ce.html http://www.cs.ox.ac.uk/people/yarin.gal/website/blog_3d801aa...