Why Google Turned Machine Learning Into Open Source
One of the fundamental traits of humans is our ability to learn through repeated exposure to external stimuli. A child, for example, can learn to distinguish between a cat and a tiger by looking at multiple images and repeatedly guessing the animals, until errors are removed, and the visual characteristics of each animal are learned. The field of machine learning tries to give computers a similar learning ability. To achieve this human-like ability, scientists have taken inspiration from the humans’ central nervous system and developed systems that are based on the somewhat similar concept of artificial neural network, in which thousands, or even millions, of computers can be interconnected like neurons. This type of agile and distributed software architecture could give computers massive computing power, but most importantly, an ability to learn and evolve, through both repeated exposure to external stimuli and pattern-recognition.
In the last four years, Google has undertaken a fundamental transition from being a mere search company to being a machine-learning organization, and has gradually turned its massive network of millions of computers into the largest artificial neural network in the world. In the process, it bought the artificial intelligence company DeepMind Technologies, whose investors included Elon Musk, and turned it into Google DeepMind, a major initiative to apply artificial intelligence and machine learning to everything that Google does. Google’s First Stage in Machine Learning
The first stage of the new artificial intelligence initiative resulted in a company-wide machine-learning system called DistBelief. This proprietary machine-learning system was used by over 50 teams at Google and resulted in a number of artificial intelligence innovations that were applied to new and existing products from self-driving cars, robots, and Internet-connected home appliances, to semantic-search algorithms in Google Search, speech-recognition in Android smartphones, spam-blocking in Gmail, visual-object recognition in YouTube, and much more.While on the surface, many of these existing products didn’t seem to change much, under the hood they gradually became significantly smarter, and evolved by self-learning the individual users’ habits and recognizing patterns.
Jeff Dean, one of the leaders of Google’s initiative explains “Machine learning is really powerful and it allows to develop various kinds of perceptual and language understanding tasks. These models allow computers to see and understand, for example, what is in an image when you are looking at it, what is it in a video clip. And that enables all kinds of powerful product features.”
The Next Frontier of Machine Learning: Open Source
Because progress in such an important area as artificial intelligence benefits from collective innovation, Google has decided to create a common standard and an open-source library for machine-learning, called TensorFlow, which will allow anybody in the world to tap into Google’s existing source-code, then build applications on top of it, and easily share the developments with the entire community.Rajat Monga, Google Technical Lead, explains that “TensorFlow is a machine-learning library that is used across Google to apply deep learning to a lot of different areas. We think that having this as an open-source tool really helps to speed up development. So we expect developers to be able to do a lot more than they can do today.”
Greg Corrado, Google Senior Research Scientist, emphasizes the importance of a common standard: “Machine learning is the secret sauce of the products of tomorrow. It no longer makes sense to have separate tools for researchers in machine learning, and for people who are developing real products. There should really be one set of tool that researchers can use to try out their crazy ideas, and if these ideas work, they can move them directly into products without having to re-write code.”
“Part of the point of TensorFlow is to allow collaboration and communication between researchers. It allows one researcher in one location to develop an idea and explore it. And then send code that someone else can use on the other side of the world.”
While Google is giving away its entire machine-learning library, it will nevertheless, greatly benefit from developers using its own standard, and from companies and universities renting out its cloud-based artificial neural network. Greg Corrado says that Google plans to cross-sell its cloud infrastructure to machine-learning developers: “We think we have the best machine learning infrastructure in the world. And we have the opportunity to share that. And that’s what we want to do here.”
