This week was so full of fun patent applications that I didn't know where to start. We've got a throwback to 2013, a virtual assistant that knows when I've stopped talking, and headphones that can determine a user's hearing abilities.
But as always, remember that the big tech companies file all kinds of crazy patents for things, and though most never amount to anything, some end up defining the future
Alphabet
Released in 2013, Google Glass, Google's stab at smart glasses, hit a fever pitch before quietly dying two years later. When it first came out, everyone thought it was the coolest thing on earth: a head-up display! On your face! That can show you your calendar! Countless articles were written about how it looked, what it did, how it worked, how it might violate privacy, and why it failed. The term Glasshole was conceived. My former colleague Kyle Russell was even assaulted while wearing a pair.
Google released version 2.0 to business users in 2019, but the initial excitement over the hardware died down and lost in the annals of tech history. Until now, it seems, at least according to this patent, which outlines a process for making Glass look better and work with glasses and prescription glasses. The original Glass had a tiny display attached to lensless glasses. But by placing a photopolymer lens in between the two pieces of glass that make up typical eyeglasses, Google hopes that the wearer doesn't look as dorky. Jury's still out on that, though.
My home is a smart home. I have a Google speaker or display in every room, I have Nest Cams, I have a Nest Hello and a Nest Learning Thermostat. I never really consider that my house is listening to me, even though I know that it is. At its simplest, I use my devices to turn on lights and play music, and at its most extreme, I use the speakers to enhance books that I'm reading to my kid, using the Read Along with Google feature. But sometimes I wonder if it's still listening even if I've stopped reading or talking to it — and especially when I've started talking about something private.
This patent takes that into consideration and offers a better way for the system to know that I've stopped talking by using various data points — such as my reading speed, other noises that imply I've moved to do something else, coughing — to signal that I've stopped reading.
Identifying explicit video content
YouTube is notorious for allowing objectionable content to seep into various parts of the app: This study shows that its algorithms often recommend false or sexualized content, explicit content is recommended to children, and YouTube still includes health misinformation. YouTube uses machine learning to try to catch some of the objectionable content before it's served up to users, and it's even hired moderators to try tackling the problem. Moderation efforts are working somewhat — in April, the company boasted that "violative view rate" was down 70% from 2017 — but clearly more work needs to be done.
This patent looks at improving how neural networks find objectionable content, using various combinations of inputs, such as comparing it to other videos that contain explicit content; analyzing the tags or title of the video; analyzing certain aspects of the video; and using various machine learning methods, such as triplet loss, to determine whether the video should be flagged.
Amazon
Hey machine, I'm talking to you
You can set up virtual assistants, like the Nest Hub and the Echo, to recognize certain voices at home. That allows the device to serve up personalized information, depending on who's asking it the questions.
This patent takes that scenario and expands on it, laying out a way for voice assistants to recognize different people outside of the home and can use natural language to give it commands. One of the examples provided imagines a coffee shop. A patron walks in, says three words to a voice assistant that's maybe installed in a POS; the machine recognizes that person's voice and pulls up account information.
Once everything is confirmed, the person could say, "I'd like a latte with an extra shot," and the system makes and then serves the coffee drink. At the same time, the computer sends a purchase order to the coffee shop's system. A hot cup of coffee without having to talk to anyone before I've had the coffee sounds like a dream.
Apple
I've been a copy editor for most of my journalism career, which means I'm expected to know how things are spelled, various grammar rules, and how to make sentences sound better. But when I write out a text or email on an iPhone, you'd think that I've never written a sentence in my entire life. My fat fingers often hit the wrong keys, and sometimes if I'm in a hurry, even my sentence structure is embarrassing.
This patent aims to help me and my fellow copy editors by teaching a machine to look out for errors, as well. Using a neural network, my phone would be able to not only correct my spelling, but also compare it to various other words that are spelled the same, and make sure I'm using the appropriate one.
This patent makes so much sense, I'm surprised it hasn't been done yet (and if it has, please email me and let me know!). Headphones as they are made now are pretty much configured for people who have no hearing loss. This patent imagines a set of headphones that can be customized to the user, by doing a hearing assessment right on the spot. After receiving the data, the headphones could automatically adjust to turn up or turn down certain frequency levels or otherwise enhance the audio to help the wearer feel comfortable.
Microsoft
Surveillance without the privacy violations
Cameras are everywhere, monitoring our every move. But when capturing this information, certain steps must be taken to ensure privacy. Blacklisting, or providing rules around what can't be seen, are often prone to errors, which can not only violate people's privacy, but can violate laws as well.
This patent lays out how to improve these methods, using whitelisting methods instead. Teaching the machine to recognize what is allowed, rather than what is not, can help improve this functionality in more-precise ways. In fact, the patent references tests where whitelisting was around 5,000 times more accurate than blacklisting.