AI is analyzing job interviews in the UK, a machine learning powered 🎨 Picasso, fiery debates on the future of AI and more — TWML
Keep your eyes peeled! One of the paragraphs in this week’s article is generated with a Transformer-type network. Can you figure out which one?
We’ve started with beer a few weeks ago. How about some pizza? Watch this robot neatly lay out (or print, if you may) a pretty good looking pie! AI? Maybe if you call expert systems and continuous control AI. Machine learning? Probably not. Nevertheless… quite cool.
Now on to something way more interesting. A new paper that came out this week by Mellor et al. combines two extremely fashionable ways to train machine learning models: generative adversarial networks and reinforcement learning. Don’t worry, we won’t talk about the math, but you do have to check out the github page accompanying the paper. It’s full of fascinating examples and video animation of how the AI agent is painting.
Unlike the previous GANs we’ve talked about in the past weeks, these agents are not generating the image pixel by pixel. They analyse the current state of the canvas and then pick an action from their “action space” defined by brush type, pressure, color, where and how the stroke starts and ends and so on. Of course, the painting is done digitally, but the same concept could apply to real-life robots… although the training process would be way slower.
An AI system is starting to be used in the UK to predict the best candidates for jobs. The AI analyzes data from video interviews (such as the sound and text transcript of what’s said and facial landmarks). The interviewing technology was developed by Hirevue, a US company.
A few weeks ago we were underlining how machine learning systems can be riddled with biases and how they can fail when they face data outside of their training distribution. This is back when we challenged you to check what a model trained on ImageNet thought abut you.
Quite a dystopian vibe, mostly due to the lack of transparency and robustness of these models. But of course, it’s easy to see the benefits if everything’s done right.
Talking of explainable and robust AI, you probably need a bit of context for the next one. Gary Marcus has recently launched his new book Rebooting AI, co-written with Ernst Davis. The book is a very welcome and well deserved critique on the current artificial intelligence state of the art, including deep learning. Needless to say that it’s been stirring quite a few [needed] debates within the AI community.
One of Gary’s main arguments is built around language comprehension. He seems to attribute this skill a fundamental role in thinking and reasoning. Of course, he is not alone in this as many other scientist have agreed on this view for the last few decades, but new research seems to suggests otherwise — that there is in fact a separation between thought and language. Check out the thread below, started by Carlos E. Perez if you’d like to read more.
We recommend reading Rebooting AI. It’s a really fresh and grounded view on current state of the art and puts recent pivotal works, such as Judea Pearl’s The Book of Why, into the bigger context of AI. If you want to hear our thoughts on the book in the form of a review, leave us a message in the comments below!
It’s that time of the year again. World Summit AI is happening next week in Amsterdam! We were there last year and we greatly enjoyed both the conference and the city. Feeling like you’re to much of an AI newbie to join? Lousy excuse! The main talks are quite noob-friendly and the networking opportunities are endless.
World Summit AI is a yearly conference aimed at the broader development of AI systems, and is a major event in the AI community. There is more than a hundred participating companies and more than 200 speakers, with more than 200 experts from academia, industry, and government.
Boston Dynamics has released a new fascinating video of their robot, Atlas, doing gymnastics. The results are so good and the movement so accurate that we suspect no reinforcement learning, or any other sort of machine learning for that matter, was used to program this.
We’ve seen some trending tweets on this stating the likes of “this is going to leave human gymnasts jobless”. This is like saying that running became obsolete with the invention of cars. It has to be one of the most absurd hyper sensationalized thing we’ve heard lately. Human gymnastics is an ode to physical excellence. Saying that having robots perform a sport will put said sport in danger is close to nonsensical.
That’s all for this week! Until next time, have a good one!
