YouTube uses mannequin challenge videos to train depth estimation models, AI essay grading systems under-deliver across the US, gastric cancer risk prediction and more — this week in machine learning.
In the spotlight 🔦 an AI powered app that helps you edit your videos 🎞️.
US is having issues with sub-par AI-powered essay graders. A report by VICE Motherboard reveals that in as much as 18 states only 5 to 20 percent of the grades are rechecked by a human teacher, most pupils being stuck with the AI assigned grade.
The system seems to be slightly overfit. All students have to do to earn a good grade is to learn and use the cues (or features) the algorithm picks up on when processing their work. Basically backwards engineering the whole AI.
“The BABEL Generator proved you can have complete incoherence, meaning one sentence had nothing to do with another,” and still receive a high mark from the algorithms.
A new research paper has been published in Nature on using machine learning techniques to predict the risk of gastric cancer based on various biological factors such as biological characteristics, Helicobacter pylori infection status, endoscopic findings and blood test results. You can find the article and the results here:
Alright, that might ended up being a bit too science-y. To counterbalance, here’s an easy to read and very down to earth article by WIRED.
It underlines the role of human beings as machine teachers and makes a philosophical argument about why machines are not smart. We, at Cognifeed, believe that comparing artificial intelligence with the natural kind is a bit misguided. As Noam Chomsky put it:
To ask if machines can think is the same as asking if submarines can swim.
Nevertheless, Ask the Know-It-Alls: How do Machines Learn? offers some pretty good insight to novices into the realm of AI and ML.
In an effort to push forward machine learning in video games, Unity 3D introduced in September of 2017 their machine learning agent toolkit. It’s a wrapper over TensorFlow that allows Unity users and game developers to dive into reinforcement learning and teach agents custom behaviors. The video below shows how Sam trained a car to park and I have to say… It seems to have learned faster than I did.
Ever thought that your YouTube suggestions are a bit… dubious? If you want to learn more about what happens behind the scenes read Forbe’s article about the ways YouTube is using AI and ML. Apart from recommendations, YouTube uses AI to detect and remove objectionable content automatically, add effects to videos and trains algorithm to predict the depth of objects (using over 2000 “mannequin challenge” videos).
Was OpenAI’s “too dangerous to release” stance on their latest natural language processing AI simply a publicity stunt? BBC took a look into the new generative technology, and their findings seem to suggest that we have nothing to worry about just yet.
Noel Sharkey, a professor of computer science at the University of Sheffield, conducted his own tests on the generator and was not too impressed.
In our opinion the relatively poor performance in fooling humans is a result of an extremely eclectic training data set combined with an architecture (transformer, but not the cool car shape-shifting ones) that’s inherently weak at dealing with long-term consistencies in text.
Startup Spotlight
I consider myself quite skilled from a digital point of view. I can use the Office suite, edit photos, write computer programs, video games and ML algorithms. Yet I’ve never managed to learn how to edit videos. I don’t know what software to use, I don’t know where and how to cut clips, I don’t know how to make transitions… I’m a video editing illiterate.
If you are like me, then TRASH is for you! Despite the name, the app seems to be quite useful for editing short video clips. It uses AI to figure out how to cut and arrange the clips on a given sound track.
I really wanted to give this a try, but as of writing this it appears that the app is an iOS exclusive.
That’s it for today! Don’t forget to follow us on twitter for more Machine Learning and Artificial intelligence news!
We’re also getting closer to our closed alpha, so make sure to enlist if you want to become a machine teacher!
Until next time!
