In our last newsletter, we mentioned academics around the world were boycotting South Korea’s KAIST University because researchers there were working on AI-powered weaponry. While the following aren’t weapons, some very interesting surveillance-related applications surfaced this past week that we think are worth the spotlight, considering how much military use of AI has been in the news recently.

But I’m behind a wall

One of the most fascinating (and potentially scary) developments in the past two weeks is a US Army project developing an artificial intelligence tool that can recognize faces through walls and in the dark. How does it do it? It uses thermal imagery scanning: the AI takes a map of your body heat and matches the facial thermal scan against a database of faces. It sounds tricky because pictures in a database are not in thermal scan form. But preliminary research has successfully matched these scans against regular pictures. So, ultimately, if you’re not on some sort of watchlist or in a database of most wanted, you don’t need to fear that this AI tool will recognize your face. Or do you . . . ?

And while we’re on the topic of surveillance . . .

We’ve heard about this type of application before: take an audio recording of a crowd of people and isolate each voice so intelligible language emerges from a garble of voices. But seeing it in action is intriguing. Google just released videos of its new voice-separation application where it shows this idea in action, albeit limited action. (It only shows two people talking at once.) In order to work well, this AI needs to see physical cues, like each speaker’s mouth moving, in order for the tool to separate the voices. It’s not a long stretch to see how this could be used for public surveillance.

Take a look at the video here:

A dangerous mix

It’s incredibly important to understand how drugs interact with each other to keep patients out of dangerous and deadly situations. Maybe seeking some positive press after their recent debacle (as mentioned above), South Korea’s KAIST University recently shared that it’s developed a deep learning system to precisely predict interactions between drugs. With an accuracy rate of 92.4, this tech can predict 192,284 drug-to-drug interactions. But it doesn’t stop there. In its most recent iteration, it can predict drug-to-food interactions as well.


Researchers use dog behavior to train neural nets
AI creates first Flintstones cartoon videos in 50 years
Machine learning improves brain-cell diagramming
Smartphone cameras use AI upgrade to reach lab-level microscopes
Artificial intelligence helps create rare “metallic glass” alloy
MIT makes road maps from aerial images