Select Page

This week in AI #6: Backdoors, new tools, fashion & much more!

by | Tech

We’re back this week with the 5 most thought-provoking articles we’ve found: two new deep-learning tools, behind the scene of Siri’s voice, backdoors in neural networks and much more! So keep on reading…

1- Brace yourself Karl!

Fashion Designer have a reputation for living in the future, but the future might already be here!

Even though there have been some interesting leads, being creative is still one task humans excel at. But for how long? Amazon just teamed-up with Whole Foods to go after your grocery shopping-list and they’re not stopping there! Now they’re setting their eyes on your fashion choices by launching Prime Wardrobe. It’s their first venture into fashion but they’re going at it with everything they’ve got and they’re already developing an AI Fashion Designer to spot forthcoming trends! Fashion has long been thought of as a sign of belonging to a certain group; maybe a day will come when all you’ll have to do is grant access to your social media and then wait for your new wardrobe to arrive!

2- Google doing its thing again

Examples of Google’s latest watermark removal on stock photographies.

Let’s say you’re in the business of selling pictures; you want your customer to be able to preview them but make sure they can’t steal them right away. The industry’s answer to this problem has been the same since the 90s: watermark. Directly overlay your brand name on top of the picture to make it unusable. But that was without counting on Google that just presented a CVPR paper highlighting how easily-removable they are. Fortunately while they dropped this bomb they also provided a defuse-kit: what to do to make their algorithm useless. Shutterstock was fast on turning this into PR gold by saying they reverse-engineered Google’s algorithm. All’s well that ends well but this really goes to show you how AI can drastically impact businesses that are taken for granted.

3- Can you hear me?

Text-to-speech synthesis process used to generate Siri’s voice.

Have you ever wondered what technology is behind all the virtual assistants that can speak? Is there someone painstakingly recording all possible words in order to later stitch them together? That’s how it used to be done but surely we can do better now. Well, wonder no more, just head over to Apple’s Machine Learning Journal and go for their On-device Deep Mixture Density Networks for Hybrid Unit Selection Synthesis or as they summarized it: Deep Learning for Siri’s Voice.

4- Backdoors everywhere!

Here a post-it is used to fool the network so that the stop sign is interpreted as a speed-limit sign.

When we hear about backdoors, most of the time we either think about some hackers working in a basement or some big government agency trying to cover its track. Well, in the case of neural networks backdoors might take on a whole new appearance. A lot of researchers in the Deep Learning field are currently working on testing the robustness of such networks and how they can be hacked. What they’re finding isn’t all that reassuring. For example, one thing that is taken for granted in the community is to use pre-trained networks and fine-tune them on the task at hand; turns out if you’re malevolent it’s not that hard to modify this pre-trained network so that it triggers your choice of specific behaviors even after it’s been fine-tuned…

5- Player 3 has entered the game

In our previous episode, we introduced two new deep-learning frameworks focusing on front-end development. It turns out those two were not the only one spending their summer polishing their latest library: we now have two other new players in the deep-learning landscape! First, Sony released their internal development tool running on Windows. Then quickly thereafter, Microsoft announced Brainwave, the deep-learning platform designed for real-time AI in the cloud. Will we end up with one framework for R&D, one for front-end development and one for production?

Let us know what you think in the comment section!

Share This