The coming year could see the rise of consumer-focused machine learning, and the beginning of the end for corporate omniscience about user data. That’s one of the implications of the latest batch of influential predictions from financial firm Deloitte, which projects that more than 300 million machine learning-tailored smartphones will be sold this year.
Right now, the most important machine learning algorithms — the ones that let us control our phones with a voice command and get predictive directions to our next destination — use mostly out-board computing power to do their jobs. They run primarily on remote data servers operated by companies like Google, and while these companies incur huge costs for running all that computation, they make it up through the increased insight into users’ lives. Sure, they have to pay for the power to parse a command to set a calendar event for 9 p.m. Eastern on Friday. But now they know you’re attending a particular local concert this week, and there’s real value to the insight that fact can provide into your buying habits.
There’s hope that any trend which moves from remote and toward local machine-learning could be a win for privacy and data security. The revolution, if indeed it does occur, would also open up the ability for much longer-term data mining from the digital lives of users. The main advantage of A.I.-focused hardware is not its speed but its power efficiency, and so it could let users simply turn on and forget a “Watch My Activity” option without dooming their smartphone to a thrice-daily charging schedule.
Deloitte is not the first to predict an A.I. revolution in smartphones, but the timeline it’s put its name to is ambitious, to say the least. If 300 million such phones are in fact sold this year, it would represent about a fifth of all smartphones projected to be sold overall — which would be a remarkable expansion.
While Deloitte makes mention of software emulators that could supplement next-gen machine learning hardware, there’s no specific mention of what newly efficient hardware or algorithms might make these predictions possible. Qualcomm has announced some chips with relatively untested machine learning architecture, but this sort of specialized hardware has not really been part of the conversation. Discussions have centered around using existing or slightly altered GPU architecture, but the efficiency gain is nowhere near new machine learning hardware (often called neuromorphic hardware).
While there are obvious security gains to keeping your info from having to travel the web, the trend toward onboard machine learning hardware could enable hackers who successfully breach a phone with the perfect onboard spy. Previously, redirecting a phone’s hardware to actively watch and analyze data on its user would be easy for security software to spot as unwanted activity — but soon, that could be a very desirable thing for your phone to be doing.