(definitely a hot take — let me know what you think!)
A few hours ago I noticed that something was bugging me about the TensorFlow announcement:
I mean, it’s yet another deep learning platform — and there are lots of them.
Those in the know — such as very smart ML engineers at Google — point out that it’s particularly close in features and approach to theano.
(FTR, I’m in that vast majority)
In other words, it’s not clear what you can do with this thing that you can’t with the many alternatives.
But I think Josh Bloom gets at the beginning of an answer:
I.e., here we have the biggest player in machine intelligence, a company that has long published papers about its sophisticated distributed systems years after their internal development but which has rarely released code for current IP, agreeing that the core software tools in this space are a commodity. Of course, we can debate whether they are leading or following on that point.
But wait, are they passing off something old and dusty?
It’s impossible to prove as an outsider, but I don’t think so. Of course, there’s plenty of exotic machine intelligence awesomeness inside the GOOG — look no further DeepMind’s deep Q learning — but I think the days are gone when they comfortably lead the rest of the industry by years in distributed software systems, for AI or otherwise. Competition is much fiercer now, and other teams have gotten very good (partly by talent cross-pollination). These aren’t the days of Map Reduce, and the mammoth of Mountain View is no longer operating in a vacuum.
And there are some very real advances in TensorFlow, from the memory optimizations to the data flow model to the slick architecture visualizations. Someone at Google is certainly working on the next thing, but this very much smells like today’s framework.
How long now before we see a hosted version on GCE, or close cousins on AWS and Azure? You can argue that there’s little truly new to see here on the technical front, and for many purposes you may be right. But my sense is that this software release marks the end of the beginning for the deep learning revolution.
Now that anyone can access the tools, the questions move to:
- Where does the data come from?
- How do you design a complete and valuable data product, of which deep learning may only be a small piece?
Google and the other internet giants clearly have good answers to these questions. But over the next few years: will other entities, big and small, bring all the necessary elements together? Or will the majority of deep learning’s value continue to be captured by today’s big players?