News Release

Artificial intelligence: Towards a better understanding of the underlying mechanisms

A multidisciplinary study from SISSA scientists employs an innovative method for studying deep neural networks

Peer-Reviewed Publication

Scuola Internazionale Superiore di Studi Avanzati

Information Processing in Deep Neural Networks

image: Deep neural networks transform the data through a sequence of stages (from left to right). view more 

Credit: Alessio Ansuini

The automatic identification of complex features in images has already become a reality thanks to artificial neural networks. Some examples of software exploiting this technique are Facebook's automatic tagging system, Google's image search engine and the animal and plant recognition system used by iNaturalist. We know that these networks are inspired by the human brain, but their working mechanism is still mysterious. New research, conducted by SISSA in association with the Technical University of Munich and published for the 33rd Annual NeurIPS Conference, proposes a new approach for studying deep neural networks and sheds new light on the image elaboration processes that these networks are able to carry out.

Similar to what happens in the visual system, neural networks used for automatic image recognition analyse the content progressively, through a chain of processing stages. However, to date, it is not completely clear which mechanisms allow deep networks to reach their extraordinary levels of accuracy.

"We have developed an innovative method to systematically measure the level of complexity of the information encoded in the various layers of a deep network - the so-called intrinsic dimension of image representations" say Davide Zoccolan and Alessandro Laio, respectively neuroscientist and physicist at SISSA. "Thanks to a multidisciplinary work that has involved the collaboration of experts in physics, neurosciences and machine learning, we have succeeded in exploiting a tool originally developed in another area to study the functioning of deep neural networks".

SISSA scientists, in association with Jakob Macke, of the Technical University of Munich, have examined how the information acquired from neural networks used for image classification is processed: "We have found that image representations undergo a progressive transformation. In the early processing stages, image information is faithfully and exhaustively represented, giving rise to rich and complex representations. In the final processing stages, the information is radically simplified, producing image representations that are supported by a few dozen parameters" explain the two scientists. "Surprisingly we found that the classification accuracy of a neural network tightly depends on its ability to simplify: the more it simplifies the information, the more accurate it is."

This is an especially important result for SISSA that has recently launched a new research program in Data Science, with the goal of studying and developing innovative algorithms for the processing of complex and large data sets.


This study has been published in the proceedings of the 33rd Annual NeurIPS (Neural Information Processing Systems) Conference, the key appointment dedicated to artificial intelligence and to machine learning, in Vancouver from 8 to 14 December 2019. On that occasion, it will be presented by Alessio Ansuini, first author of the study and the scientist who was responsible for conducting the experiments during his post-doctoral research activity at SISSA.

Disclaimer: AAAS and EurekAlert! are not responsible for the accuracy of news releases posted to EurekAlert! by contributing institutions or for the use of any information through the EurekAlert system.