Tags
Deep Learning: Who are the active users? What problems are they trying to solve?
In 2013, Google purchased a firm called DeepMind Technologies. DeepMind is an artificial intelligence firm specializing in deep learning techniques. The acquisition of DeepMind is an outgrowth of Andrew Ng’s Google Brain. Google Brain is described as a complex neural network overlaying a set of deep learning algorithms. It has demonstrated that it can categorize images by their subject (e.g. “cat”) without ever being given a list of properties of that subject. Google expects to use DeepMind’s human resources and experts to improve their own image search and develop new, sophisticated forms of unsupervised learning. [1]
Unsupervised learning is DeepMind’s specialty and a hallmark of deep learning. One article notes that DeepMind developed software that learned how to play seven different Atari video games using only the games’ user interface. [2] Interestingly, unlike most game-playing systems, this project did not use an internal data stream – it took the frame-by-frame visual output of the Atari and converted it into a stream of 84 x 84 pixel images and using this as inputs to train a neural network to attain high scores – but without providing the system any specific knowledge of the rules of the game or particular strategies for maximizing scores. [3]
While playing an Atari game may seem like a relatively trivial task, it underscores the key elements that set deep learning apart from other data science tools. First, it takes as input a relatively complex data type: Images, video, and natural language are typical input types for deep learning applications. Second, it employs unsupervised learning to develop a neural network able to distill this complex data type into a single abstract output, be it a category (cat vs. dog or optimal play strategy).
GigaOm provides an intuitive explanation of the key defining feature of Deep Learning: “It’s about teaching machines to think more hierarchically or more contextually – To see a picture of a mole, for example, and work down from recognizing the features that comprise an animal to recognizing the specific features that make it a mole.”[4]
GigaOm provides several other examples of private companies developing their businesses around Deep Learning. Cortica, for example, is commercializing image processing and recognition by offering a system that enables advertisers to serve content related to web pages without the use of embedded image metadata.[4] This software was borne out of research done by Igal Raichelgauz, CEO and co-founder of Cortica, on actual human brain tissue. Cortica claims that their software, like most commercial applications of deep learning, functions in ways similar to the human brain [4], e.g. according to a hierarchical classification scheme in which the learner increasingly refines its categorization of an image.
[1]http://www.fastcolabs.com/3026423/why-google-is-investing-in-deep-learning
[2] http://www.technologyreview.com/news/524026/is-google-cornering-the-market-on-deep-learning/
[3] https://medium.com/the-physics-arxiv-blog/bfc25f2ffe03
[4] http://gigaom.com/2013/11/01/the-gigaom-guide-to-deep-learning-whos-doing-it-and-why-it-matters/