Microsoft is developing an AI called “Adam”, which is principally from the same family as Google’s artificial brain.
All the big tech companies, drawing from a body of academic research, are pioneering a form of AI called “deep learning”. This is the AI that’s behind speech recognition tech like Siri and voice search on Android phones, reverse image search, and language translation. Thus far, these technologies hold records for accuracy in those regards.
Adam, Microsoft claims, has beaten those records. Adam is twice as adept at recognising images while using 30 times fewer machines. Peter Lee, head of Microsoft research, boasts that Adam tops the Google Brain’s performance on a benchmark test called ImageNet 22k. The test is rather self-explanatory: how well does your AI process a database of 22,000 images?
However, Adam isn’t gunning to replace Google Brain outright. It’s really about optimising the system, how it handles its data and how the servers communicate with each other. Any system descended from deep learning is about modelling the way the human brain works, specifically the networks of neurons, and then processing data the way our brain does.
Of course, as we scale these systems to greater heights, it gets more and more unwieldy to replicate. Adam solves this problem by asynchrony, which basically splits the datastream to be solved by separate systems, and then merges the end result together. This ends up being less problematic than it sounds.
Take, for example, a picture of your dinner from last night. One server tags it “dim sum”, another goes “dumplings”, yet another goes “Chinese food”. Yup, science has finally automated your friend that dumps 20 similar hashtags on their Instagrams. (Good, they can finally be systematically removed from all society.)
Microsoft claims the project is still embryonic, and so far no plans to release it to the public have been made. It will, however, find its way into current Microsoft products such as Bing.