Now that you’ve transformed your system through successive evolutions of architecture goodness...you've made it cloud native, you now treat a fist full of datacenters as a single computer, you’ve microservicized it, you’ve containerized it, you’re continuously releasing and improving it, you’ve made it reactive, you’ve socialized it, you’ve mobilized it, you’ve Hadoop’ed it, you’ve made it DevOps friendly, and you have real-time dashboards that would make NORAD jealous...what’s next?
Deep learning is what’s next. Making machines that learn. The problem is how?
All the other transformations have been changes good programmers can learn to do. Deep learning is still deep magic. We are waiting for the Hadoop of deep learning to be built.
Until then, if you aren’t Google with Google sized clusters and cloisters of PhDs, what can you do? Greg Corrado, Senior Research Scientist at Google, gave a great presentation at the RE.WORK Deep Learning Summit 2015 (videos) that has some useful suggestions:
-
You can do a lot of experimentation with very little money. You don’t need a huge infrastructure team or a lot of investment.
-
Even if you don’t have deep pockets you probably have a desktop computer. That computer has a GPU in it that you can upgrade at a nominal cost.
-
A desktop computer is sufficient to process 100 million images, for example. You need some really good GPUs and people who know to program them. As for many early technologies programming prowess is often fungible with expensive resources.
-
You can experiment and explore these ideas with existing open libraries: Caffe, Torch, Theano.
-
Greg hopes in the future there will be services available, but currently you are mostly rolling your own.
-
If you are building a technology company today you should expect to use supervised learning (learn by example).