Scientists from the Department of Computer Science at UTSA, have managed to change the conventional approach used to explain the different decisions of machine learning that is based on a single injection of noise in the input layer of a neural network.
Experts have shown that adding noise, also known as pixelation, across many layers of a network would provide a much more robust representation of an image recognized by artificial intelligence (AI), creating more robust explanations for decisions made by is.
This work would be useful in helping to develop what is currently known as “explainable AI” , which seeks to enable high-security artificial intelligence applications, such as autonomous driving or medical imaging.
As Sumit Jha, one of those responsible for the study, points out, “it is about injecting noise into each layer”. Because of this, “the network is now forced to learn a much more robust representation of the input in all its internal layers. If each layer experiences more disturbances in each of the trainings, then the representation of the image will be more robust and you will not see that the AI fails simply because it changes some pixels of the input image.
The work carried out by the team from the University of Central Florida, the Air Force Research Laboratory, and SRI International is a significant advance over previous work carried out in this field. Already in 2019 they demonstrated how the bad conditions present in nature could lead to a dangerous neural network.
At the time, the researchers asked a computer vision system to recognize a vehicle located on a road, and it did so correctly. The team then added a small amount of fog, raising the same question to the network again. Artificial intelligence, however, identified it as a source.
Computer vision , which is understood as the ability to recognize images, has many commercial applications, since it allows, for example, to better identify some areas in the liver and brain of cancer patients, which tend to be of considerable concern to patients. the doctors.
But this type of machine learning can also be used in many other industries. Drones can use them to help detect leaks in pipes, while farmers may already be using them to detect the first signs of crop diseases, thus improving their yields.
On the other hand, through deep learning , a computer is trained to carry out different behaviors, such as making predictions, recognizing speech or even identifying images. Rather than organizing the data for the execution of established equations, deep learning tends to work within a series of basic parameters on a given data set, training the computer to learn on its own by recognizing a series of patterns using many processing layers.
