Before winning the Nobel Prize in 1906, Santiago Ramon y Cajal He wrote the doctrine of neurons. He described them as individual units that communicate with each other in a directional way through the space between them. Thus, he outlined in general terms how information flows in the brain. For this reason, his discoveries not only form the basis of modern neuroscience, but also that of the development of artificial intelligence (AI).
This is the case of the neural networks: mathematical models that teach computers to process data as a person would. Their rapid evolution in recent years has allowed them to be used to treat complex problems due to their versatility and power. Thus, they have come to permeate many areas of daily life, from social networks or mobile suggestions on how to improve digital health to the famous ChatGPT. This also seems to be the case in industry, although companies often run into some difficulties in their implementation.
“The functioning of deep neural networks depends on their ability to merge data appropriately for the identification of the most relevant features. He The problem is that most of the networks available on the market implement the same fusion techniques, which reduces their flexibility when adapting to specific problems. In addition, to improve the results obtained, traditionally the number of network parameters is increased, which makes its use slower and more expensive,” he points out. Javier Fernandezresearcher in the Artificial Intelligence and Approximate Reasoning Group (GIARA) of the Public University of Navarre (UPNA).
Those who wish to know more about the project can do so through this link
Hence, the educational center began to explore new algorithms to respond to the challenges of companies in this field. A job in which he has established synergies with NAITEC. «In our daily contact with the industry, we identify a series of needs, in terms of processes and products, that require artificial intelligence. B.We look for them to have more associated services through these techniques developed by the UPNA”, indicates Javier Ojer, Head of the Mechatronics Business Unit at the technology center.
For this reason, both entities continue their collaboration through the project fusiprod. Its objective? Improve the information fusion processes within the network, so that its performance can be optimized at a lower cost. The initiative is coordinated by ADItechin turn coordinating agent of the Navarro R+D+i System (SINAI)and financed by the foral executive in the call for aid to technology centers and research organizations to carry out collaborative R&D projects.
TOWARDS A “MORE INTERPRETABLE” AI
After being trained, neural networks are able to predict the tastes of a person or when a windmill is going to break down. For this they need, on the one hand, the data and, on the other, a way to extract the information. For this reason, the first stage of the project focused on data collection, a task carried out by NAITEC.
«Fusiprod focused on a specific case: the prediction of anomalies in machines, that is, events that are not part of their activity history. The information we obtained came from time series in our test bench. In this part of the project, the collaboration was close because we had to ensure the quality of said data and that they could be used in the theoretical developments of the UPNA”, points out Ojer.
In parallel, the university group focused on “getting into the guts” of neural networks to modify them through f-based techniques.aggregation functions and fuzzy integrals. “The former make it easier to gather data and represent it well. The latter allow take into account how the data is related, if some relationships are more important than others,” says Fernández. So, rather than increasing parameters to improve results, the team sought to fine-tune the ability of neural networks to identify which data to merge.
In this sense, the initiative represents another step on the road to developing “more explainable and interpretable” artificial intelligence systems, in the opinion of the UPNA researcher: “The problem with neural networks is that we input data, but we don’t know why.” which give a concrete answer. By modifying their interior, we can try to understand what is happening inside when they operate. There are problems where it is not important to know the reason for the answers, but it is interesting for more sensitive applications such as medical ones”.
APPLICATION IN THE INDUSTRY
Finally, NAITEC was in charge of embedding the neural networks developed by the UPNA in different hardware in order to integrate the devices into products or at machine level “at a low cost”, specifies their Head of the Mechatronics Business Unit. So the main challenge was reduce the complexity of the model so that it would fit in the memory of a microprocessor and give the same results: “HWe have managed to integrate the networks developed at Fusiprod so that they have even more intelligence functions, thus increasing the competitiveness of the products.”
The results of the initiative are already being used in innovative projects with which NAITEC supports companies and entities. TOFurthermore, another of the advantages of the device is that it does not require a connection to the cloud. «There are companies that are jealous of their know howso they are reluctant when it comes to taking sensitive information from their processes to an external server”, concludes Ojer, who highlights the path that the project can still have in the productive fabric.
Enter here to read more about innovation in Navarra.