Artificial Intelligence and Machine Learning usually work best with a lot of horsepower behind them to crunch the data, compute possibilities and instantly come up with better solutions.
That’s why most AI systems rely on local sensors to gather input, while more powerful hardware in the cloud manages all the heavy lifting of output. It’s how Apple’s Siri and Amazon Alexa work, and how IBM Watson can tackle virtually any major task. It is, though, a limiting approach when it comes to making smarter Internet of Things and applying intelligence when there isn’t Internet connectivity.
“The dominant paradigm is that these [sensor] devices are dumb,” said senior researcher with Microsoft Research India, Manik Varma.
Now, Varma’s team in India and Microsoft researchers in Redmond, Washington, (the entire project is led by lead researcher Ofer Dekel) have figured out how to compress neural networks, the synapses of Machine Learning, down from 32 bits to, sometimes, a single bit and run them on a $10 Raspberry Pi, a low-powered, credit-card-sized computer with a handful of ports and no screen. It’s really just an open-source motherboard that can be deployed anywhere. The company announced the research in a blog post on Thursday.
Microsoft’s work is part of a growing trend of moving Machine Learning closer to devices and end users.
Earlier this month at is annual World Wide Developer’s Conference, Apple announced new Machine Learning APIs (Vision and Natural Language) that allow developers to add machine learning-based intelligence to their apps with just a couple of lines of code. They also unveiled Core ML for developers more well-versed in AI to take full advantage of all inference capabilities available on the local hardware. Apple’s model does have the developers train their Machine Learning algorithms on libraries Apple provides. The system then converts the code to run the AI locally.
Obviously, in Apple’s case, that hardware is inside a $700 iPhone and the CPU is much, much more powerful than anything found on a Raspberry Pi. Still, the trend is clear. These companies are moving intelligence closer to the local hardware and, where possible, relying less on constant access to massive data and intelligence stores in the cloud.
“If you’re driving on a highway and there isn’t connectivity there, you don’t want the [AI] implant to stop working,” said Varma in the blog post. “In fact, that’s where you really need it the most.”
It’s an approach that will make sense for smaller, sensor-based tasks that can learn by location, intention, recent action and the device data. In the near term, it won’t be a solution for, say, coming up with new cancer therapies (one of the areas of interest for IBM’s Watson AI).
As for Microsoft, this Raspberry Pi breakthrough is simply phase one in a quest to compress neural networks so much that they can run on a breadcrumb-sized micro controller. To get there, the machine learning models need to be, according to Microsoft, as much as 10,000 times smaller. That’s a problem the team is still working on.
In the meantime, Microsoft released previews of the Raspberry Pi-sized machine learning and training algorithms on GitHub where enterprising developers can try them out and, potentially deploy on Raspberry Pi 3 and Raspberry Pi Zero.
Ultimately, this is another piece of Microsoft’s growing Intelligent Edge strategy, which Microsoft CEO Satya Nadella outlined earlier this year a Microsoft Build developers conference. Microsoft hopes to see these tiny AI-able microprocessors deployed in everything from our offices to the clothes we wear.
For Varma, who is visually impaired, the research is a little more personal. His team is already developing a prototype intelligent walking stick to showcase their research.