Tech companies created the beast, but even they don’t know how to fully explain it. The role of trying to understand how AI decision-making works is left to mathematicians, who are in search of “explainable AI”. They are getting closer and closer.
Mathematics can be used to understand AI choices, scientists seeking “explainable AI” believe. These mathematicians try to understand it by detecting patterns.
“We created neural networks that can perform specific tasks, at the same time allowing us to solve the equations that govern the activity of the networks”, he says Lyle Mullerdirector of the Fields Lab for Network Science, at . “This mathematical solution allows us to ‘open the black box’ to understand exactly how the network does what it does.”
Starting from basic shapes such as squares and triangles, the team responsible for the PNAS in November created a recurrent neural network (RNN) capable of identifying AI components. This allows you to create dynamic, fluid patterns in an image.
It was, therefore, possible detect patterns even, notably, in a complex image of an animal in the jungle.
The RNN works without training, using a set of predefined parameters to identify objects from simple geometric patterns to natural scenes.
The network, called complex-valued RNN (cv-RNN)is based on the parameters of amplitude and phase (each pixel is represented as a complex number) and recurring connectivity (a pixel influences neighboring pixels).
With this methodology, the team even managed to connect your network to a living brain cell.
“This kind of fundamental understanding is crucial as we continue to develop more sophisticated AI systems that we can trust,” said Roberto Budzinski, a research team.
“This is just the beginning,” concludes Muller. “We believe this mathematical understanding can be useful far beyond this first example.”
