Open AI and Google have created a service allowing to observe neural networks under a microscope while IBM trains its AI on photos without owners’ permission. Follow our weekly digest for more news about AI.
Japanese startup Vaak launches the sale of an AI identifying shoplifters
Japanese startup Vaak has developed an AI-enabled system that would help prevent shoplifting. Artificial intelligence uses recordings of surveillance cameras to identify a potential thief by gestures and mimics denoting fuss and anxiety.
Vaak’s system has already been tested. In 2018, it identified a shoplifter at one of the stores based in Yokohama. In spite of the early stages of its development, the algorithm managed to recognize a criminal overlooked by guards. In the end, a thief was caught.
This week, Vaak has launched the official sale of its service.
IBM trains AI on photos without owners’ permission
IBM has created a database of Flickr’s photos in order to train a face recognition system without the permission of its authors and people in the pictures.
Flickr’s owner Yahoo handed over the photos to IBM. It collected a database with 99 m images for research purposes and obtained permission. However, the license doesn’t mention the use of people’s photos to train AI.
IBM representatives claim that if some user wants to delete images from the database, they just need to send their links to the company’s employees. However, it’s now impossible because the organization has closed access to the database. Therefore, users cannot see which photos are used for sure.
Microsoft’s new app to let the blind see images on a smartphone
Microsoft has released a new Seeing AI’s version that would use computer vision to help blind and visually impaired people see images on smartphones or PCs.
To launch the app, a user has to touch a smartphone display. The computer vision system will recognize image details and read them out through a speaker or headphones.
Free download of the app on App Store and Google Play became available on March 13.
Neural network at the Picasso’s price: a new AI-produced painting to be sold at auction
A neural network capable of generating unique people’s portraits was presented at London-based Sotheby’s auction house on March 6. ‘Memories of passers-by 1’ will become the second AI-made work-of-art sold at such an event following ‘Portrait of Edmond Belamy’ auctioned off at $432 000 at Christie’s.
As commented by the creator of the project Mario Klingemann, the buyer of this lot will provide original portraits of a lifetime since the neural network produces unique paintings.
Open AI and Google present a service capable of seeing through networks – Activation Atlases
Open AI and Google have created a tool for the visualization of interactions between neurons in AI-powered systems called Activation Atlases. This service helps understand AI-fueled decision-making process.
Earlier, there were attempts to develop the same solution but they looked like searching for separate letters in a visual alphabet of the algorithms. The new research performs the function of a vocabulary since it explains how the neural network combines these letters into words.
As reported by Open AI, this technology would prevent unexpected problems in neural networks, for example, in the event if the network uses false correlations when classifying images.
You'll know more than your colleagues and business rivals.