The first AI-powered heist on record 💰, Facebook uses Minecraft in efforts to train a more general AI, human annotators dealing with disturbing tasks and more — This Week in Machine Learning.
In what appears to be the first AI money heist ever, a company’s employee was tricked to wire $240,000 to an offsite account. How, you ask? With a voice deepfake. A voice reproduction machine learning model was trained on sound clips of the company’s CEO speaking and then used to impersonate him.
The technology is nothing new, but it got uncannily good over the past year. Recall for a second how natural sounding the voice of the google assistant is (yes, that voice is borrowed too). One start-up focusing on “voice stealing technology” is Lyrebird and other open source alternatives also exist, such as Deep Voice. But don’t panic just yet — the tech is not perfect. How can we tell fake voices from real ones? Well… with more AI, of course. And mice.
Only two weeks ago we had Generative Adversarial Networks (GAN) in our weekly digest. Now GANs are in the news again and they keep getting better. There’s a lot of research that tackles one of the core problems with generative AI. The things they generate are very… hit or miss. When we need a picture created we know, to some extent, what we want out of it.
Károly Zsolnai-Fehér, from Two Minute Papers, has made a video review on a recent paper that studies the latent space of vector that GANs base their generation from. Identifying correlations between certain manipulations of these vectors and the features of the generated image is the key to controlling a GAN’s output.
Ever wanted to see Charlie Chaplin in color? Now you can, thanks to machine learning. ML models have become quite good at video colorization tasks, but up until now the output was riddled with artifacts and the temporal cross-frame consistency was a mess. A new paper from CVPR2019 sets to fix this!
If you want to learn more about the technology, Synced have a pretty good medium article on it:
Games, video games and game engines have always been great training grounds for AI, mostly in the field of reinforcement learning. It’s no surprise that Facebook decided to use Microsoft’s Minecraft as a playground for their experiments with AI agents. Given the creative potential a world like Minecraft has, Facebook hopes it will be able to train a more general sort of intelligence than what we’ve seen so far (trained on games such as Go or StarCraft).
One thing is clear. The set of possible actions and goals an agent can have in such a world is virtually boundless. Hopefully the AI won’t end up lying belly up in a virtual blocky bed pinned down by extreme existential dread.
As you might now, supervised learning techniques require large amounts of annotated data. More often than not researchers and companies choose to outsource data collection and its annotation. The most popular way of doing so is using Mechanical Turk, a crowd sourcing platform by Amazon.
The majority of the crowd completing tasks is from underdeveloped or 3rd world countries. This fact encourages predatory tactics and unethical pay form the companies using the services. Most tasks have a reward of just a few cents.
And if that wasn’t enough, the whole ordeal is also taking a psychological toll on MTurkers. An investigation by Irish Times shows that the human annotators have to deal with a great deal of disturbing images — from colonoscopies to porn.
A new dataset of images aims to enable machine learning to estimate emergency situations. The dataset, put together by researcher from MIT, contains both aerial and ground level pictures of natural disasters and man-made accidents. The article title, while humorous, is quite click bait-y. If fed into a model trained on images of cats and dogs the same image would probably be labeled as a cat.
AI thinks this flood photo is a toilet. Fixing that could improve disaster response.
Startup Spotlight
Mining is unquestionably harmful for the environment, but it is (and will remain for the foreseeable future) crucial to technological advancements, infrastructure and economy in general. Of course, living in a world where everything can be recycled is something to aspire to, but until then we could start by minimizing the impact mining activities have on the ecosystem. This is exactly what Earth AI sets to do.
Earth AI uses machine learning and artificial intelligence in three key areas: mineral exploration, geo-surveys and autonomous drilling (which they claim to have a very low footprint). Their vision is quite inspirational — although they are starting here, on Earth, they are aiming for the stars.
