Inside a soundproof box is one of the worst neural networks in the world. When it was presented with a picture of a handwritten number (six), it spat out the result after a moment: zero. Cornell University physicist and engineer Peter McMahon, who is leading the project, defends the AI with an embarrassed smile – after all, the number was really sloppily written down.
Despite the poor demonstration, the neural network is a sensation. Opening the box reveals no computer chips, but a microphone aimed at a titanium plate with a speaker behind it. While ordinary AI works with zeros and ones, this device uses continuous sound waves. It converts the pixels of an image into sounds. Even its developers can hardly believe that the machine is sometimes right. Despite their primitive capabilities, McMahon and his team hope such sound-powered systems will revolutionize computing.
When it comes to machine learning, the rule of thumb usually applies: the more, the merrier. If you equip a neural network with additional artificial neurons, it will be better able to distinguish a dachshund from a Dalmatian. In the meantime, the programs even master creative tasks; they write essays or produce pictures. This potential has motivated many researchers to look for more efficient calculation methods.