Deep learning seems very inaccessible compared to other machine learning techniques. I'm predicting that within a year or two DL with be just another tool in the toolkit of data scientists, no longer seeming so "exotic". The nolearn library looks like a nice step in that direction, especially as it has the same API as scikit-learn.
I think it might take longer than a year or two, but it's definitely getting closer. The real issue is how fast (and easy) GPU support progresses. The CUDA Drivers and SDK make it substantially easier, but it's extremely non-trivial to train networks for the PASCAL, CIFAR, or ImageNet challenges due to the fact that there is just so much data. Furthermore, Conv. Nets still have a ways to go before they are ready to go, out of the box, similar to the functions already implemented in scikit-learn.
(Author of deeplearning4j here) Let me add a counter point to GPUs. While they are great for scaling ( I have them in deeplearning4j myself) they work great for larger problems, but you will not see insane performance boosts on smaller problems.
This is why I have a multi threaded/distributed mode for people. Sending data to the GPU is still a huge bottleneck. While great for harder problems, the problems sklearn solves are not the same kind deep learning will be used for.
Deep Learning also requires a lot of visual debugging tools very specific to neural nets. You need to be able to adjust the pace of learning as well as knowing what kinds of nets apply to what problems. This is very hard to put in to a general purpose framework.
I also don't think the GPU solution is the best for more enterprise solutions. If an enterprise company is interested in computing, they likely have a cluster of same machines. And if they are interested in data processing, the cluster is likely running Hadoop.
To that end, it also makes sense to start developing Deep Learning for Hadoop based systems to utilize the architecture that already exists in the enterprise space.
Right. Which is why I supported CPUs first with the ability to plug in different backends. Blas is supported in spark and CDH leading the way for deep learning solutions for hadoop. I hope to lead that front with what we have going on now.
