Researchers at São Paulo State University (UNESP) in Bauru (São Paulo State, Brazil), in collaboration with colleagues at Friedrich-Alexander University of Erlangen-Nuremberg (FAU) in Germany, are planning to optimize advanced artificial intelligence (AI) techniques that enable computer algorithms to collect and interpret big data as a basis for making predictions and generalizations.
The project, involving André Carlos Ponce de Leon Ferreira de Carvalho, a professor at the University of São Paulo’s Mathematics & Computer Science Institute (ICMC-USP) in São Carlos (São Paulo State, Brazil), and André Fujita, a professor at the same university’s Mathematics & Statistics Institute (IME–USP), was selected in the third call for proposals by the 2016 São Paulo Researchers in International Collaboration (SPRINT) program and announced in January 2017.
The aim of SPRINT is to promote the advancement of scientific research through collaborative medium- and long-term projects by researchers affiliated with higher education and research institutions in São Paulo State and scientists at similar institutions abroad.
A record 24 proposals were selected in the third call. These included researchers from São Paulo State and seven higher education and research institutions in five countries with which FAPESP has cooperation agreements. Additionally, four proposals were selected from researchers whose partners are affiliated with institutions in four different countries with which FAPESP does not have current agreements.
A SPRINT call will remain open until April 24, accepting the submission of new mobility proposals involving researchers in São Paulo State and 14 higher education and research institutions in nine countries with which FAPESP has cooperation agreements.
“This is the third project that we’ve submitted, and we have now been selected to participate in the SPRINT program,” said João Paulo Papa, a professor in the Computer Science Department at UNESP in Bauru and the researcher responsible for the project on the Brazilian side.
“We conducted a project in 2015 in collaboration with colleagues at Ohio State University in the US and RMIT University in Australia, also in advanced AI optimization techniques,” Papa told Agência FAPESP. “In the latter case, the focus was on image diagnosis of diabetic retinopathy [damage to blood vessels of the retina tissue caused by diabetes], and now we’ll use the same approach in bioinformatics [the application of computational techniques to the analysis and modeling of data obtained from biological research].”
According to Papa, deep learning is among the most advanced techniques currently used in AI to analyze and to extract knowledge from large datasets. Deep learning algorithms are already used by companies such as Facebook to recognize photographs and other images.
One of the advantages of deep-learning algorithms is their ability to perform unsupervised feature learning using unannotated data.
To perform such tasks, however, these algorithms must process hundreds of parameters, Papa explained. “The challenge of creating deep-learning algorithms is precisely that of choosing the most suitable parameters, because each application requires a different configuration,” he said.
The approaches used to identify the best parameters for a given application include running an algorithm a large number of times and selecting the best result.
To reduce the time required for the parameter selection process, the researchers plan to investigate the use of bioinspired algorithms, so called because they are inspired by nature. One of these bioinspired algorithms, for example, is based on the behavior of ants.
Ants tend to choose the shortest route to wherever they want to go because as they move they release pheromones to enable other members of the colony to follow them. If they chose a long route, the pheromones would dissipate, and the other ants would lose the trail, Papa explained.
“Bioinspired algorithms based on ant behavior use this premise to select the best parameter values for a specific application in a reasonable timeframe to optimize or to reduce the error rate in an AI application,” he said.
The researchers chose bioinformatics to validate the use of these techniques because their collaborators at the University of São Paulo and Germany’s FAU were interested in bioinformatics.
For example, one application they are considering involves comparing the DNA of two people to determine their level of similarity.
“The aim of our project is to use a concrete example from bioinformatics to show that machine learning and optimization methods can be used in synergy,” said Alexander Martin, a professor in FAU’s Mathematics Department and principal investigator abroad for the project.
In his opinion, FAPESP’s SPRINT program will fund the exploratory phase of the project, a phase that is often difficult to fund through programs run by other research-funding agencies in Brazil and Europe.
“SPRINT will allow us to pursue this line of research in collaboration with our colleagues in São Paulo,” he said.