Quantum computing is coming and will level up AI ⬆
And how to measure emissions from training neural networks 😷
|jackie snow||Oct 25, 2019|| 1|
Hi! This title idea came from a story I loved a few years ago, arguing that the Anthropocene, defined by humans, is just about over and the machine era is coming. “Machinocene” hasn’t been picked up yet, but maybe I can make it happen? Maybe with a tagline for easy reference? Hmmmmm?
Story of the week
Google claims that they achieved “quantum supremacy,” a very complicated process that I will oversimplify as a way to do computing in a whole new way that isn’t possible with the 0s and 1s in traditional computing. Quantum computing would be extremely fast and powerful and unlock new digital abilities, some potentially good, others potentially very bad. Our AI processes would be smarter! Encryption could be broken!
It’s a big deal, if it actually happened. There is a lot more than just bragging rights and money at stake: countries are treating it as a national security concern. But IBM and others have pushed back on whether Google actually needed quantum computing to do the process outlined in the paper Google pubbed. Even if this isn’t the week quantum supremacy happened, plenty of researchers are trying to be the first one to the finish line.
Defeating face rec tech, one (or two) stickers at a time.
Sign of the times: The London police shared training videos of terrorist attacks with Facebook to help train the algorithms to detect live-streamed shootings.
A Russian activist sued the Moscow city government over their face recognition technology that she claims violates privacy.
Surveillance in school is increasing over fears of shootings. Too bad the tech isn’t proven yet and privacy is being ignored.
Axon, best known for the Taser, plans to add license plate recognition to their police dashcams offerings and may or may not take into account the company’s ethics board’s concerns.
The robotic hand story from last week is causing some infighting in the AI community, with some arguing the results were overstated.
Beyond offensive keywords, there are also some *issues* with the photos in ImageNet.
Semantic Scholar uses AI to organize and uncover journal articles. The natural language processing tech recently expanded from a handful of topics to nearly all areas of science and now has taken a look at over 175 million papers.
Homeland security put out a request for proposals on face rec tech for the border.
Forget Big Macs, it’s all about Big Tech at McDonald’s now.
Another day, another way to identify you: AI can match a photo of someone to their MRI scan.
An algorithm that has been used on millions of patients to predict who will benefit from extra medical care is biased against the health needs of the sickest black patients.
Google trained a neural network to predict the smell of a molecule based on the chemical structure.
A startup called Groq wants to take on the biggest companies in the biz and build a new type of chip for AI applications.
Who will watch the watchers (and keep a lookout for bias)? Increasingly, people are launching startups with the purpose to keep an eye on AI.
Recently, researchers have started pointing out that fancy neural networks are very energy-intensive. One study found that training one could release five times the lifetime emissions of a car. Some algorithms are worse than others, but there are also considerations like what energy is powering the computers doing the training.
Now, researchers out of Canada have released a paper documenting a calculator they say can estimate an AI’s emissions. The calculator takes into account what cloud service is being used, where those servers are physically located, how much training time is needed, and the hardware doing the work. Obviously, this isn’t too detailed and misses plenty of possible setups, but it is still information for environmentally conscious researchers that are mindful of their impact. Even doing something like switching from servers in Australia (which rely on more carbon-intensive power sources) or change out more efficient hardware could make a difference.
Datapoint of the week
10,561: The number of missing children photos that were match with other shots from the country's orphanage institutions using face recognition, according to CNN Business.
“Glue some googly eyes to your webcam. It’ll make it easier to maintain eye contact.”
- A Reddit user on how to do well on AI-driven interview, according to The Washington Post investigation
Until we all have our own peace robots (and Nobel Prizes),