Under the leadership of the ELKH Alfréd Rényi Institute of Mathematics (Rényi), a large-scale National Excellence Program targeted at researching the mathematical foundations of artificial intelligence has been implemented. The program funded by the National Research, Development and Innovation Office (NKFIH) has achieved a number of outstanding results in terms of science, infrastructure development, education and innovation. The last ten years have seen revolutionary breakthroughs in artificial intelligence around the world. By funding the project, the NKFIH aimed to support Hungary's efforts to regain lost ground in the field of machine learning, as it is of strategic importance that Hungary does not miss out on the economic, social and scientific benefits of these innovative results. The other main objective of the project was to create and develop knowledge centers where researchers from different areas of basic research are involved in artificial intelligence research.
In addition to Rényi, the ELKH Institute for Computer Science and Control (SZTAKI), Eötvös Loránd University (ELTE), Pázmány Péter Catholic University (PPKE) and the University of Szeged (SZTE) also played important roles in the implementation of the program's objectives. The main results of the project were presented at a closing conference on 17 September 2021.
At the time of the Covid outbreak, the program was half-way to completion. This could have led to the research being interrupted, but just the opposite happened. In a difficult period, all participating institutions threw themselves into the project with renewed vigour and contributed to the results of the research teams already working on the topics in question. Significant progress has been made in representation learning, interpretability and statistical machine learning, as well as in several other areas.
Main research areas and findings of the consortium members
At Rényi, the project placed particular emphasis on research into the theoretical foundations of artificial intelligence. Of particular interest are, among other areas, the results of their research on the mathematical properties of information metrics and the detailed description of the structure of compressed information that can be extracted from large data sets using higher order Fourier analysis. They developed novel methods for comparing the internal states of artificial neural networks and have used this toolkit to investigate the properties of the learning process. They have also developed models for the control of automatic verification systems for theorems based on reinforcement learning. Based on these results, they have conducted combinatorial game theory research, investigated the smoothness behaviour of mappings implemented by artificial neural networks, and described loss function members that promote smoothness.
The research work carried out in a cooperation between Rényi and MedInnoScan Ltd focused on the practical application of artificial intelligence. As part of the program, a medical pilot application was developed to improve the treatment of around 200,000 patients with chronic wounds in Hungary. The treatment of these patients places a significant burden on the entire health care system, as they require continuous treatment, with dressings every two days on average. It makes sense to change the type of dressing depending on how the wound changes, and the artificial intelligence application embedded in the mobile app, the prototype of which was created by the researchers and presented at the closing conference, will help in this decision. The image database on which the research and AI training is based has been compiled through a national partnership. Led by the University of Pécs and with the active participation of nearly 70 medical institutions, four medical universities, the National Medical Rehabilitation Institute, a dozen hospitals and more than 50 nursing services and private wound care centers, 220,000 images of the chronic wounds of 5,500 patients were recorded. After clinical trials, the product is likely to be made available to professional nurses next year.
The main task of SZTAKI under the program was to carry out basic research motivated by practical needs, and to apply and demonstrate the results. Their research covered the information geometry foundations of deep learning, network theory, recommendation systems and reinforcement learning, and their applications spanned areas such as machine perception, social media analysis, robotics and autonomous transportation. During the three years of the program, a number of outstanding scientific papers were published and their results were presented at several public events. Key demonstrations: a portrait drawing robot; robotic applications, control and manufacture of drone flocks with reinforcement learning; a solution for exploring the inner functions of recommendation systems; detection of certain traffic objects; and an opinion analysis system for Covid vaccinations.
The ELTE Institute of Mathematics studied several areas of machine learning and their use in applied mathematics. In machine vision, the Institute investigated the teaching of deep neural networks and its specificities in several application domains (e.g. in the processing of medical images and satellite images). In the application of mathematical modelling, the efficiency of algorithms and procedures already used in practice was investigated. One such topic is the use of neural networks to support holistic programming technologies that play a central role in solving logistics or optimization problems in general. Another solution is to use neural networks to estimate parameters of stochastic processes, which are also used in the analysis of financial processes. In the context of the previous topics, research has extended to the implementation of deep learning models of natural language processing in other, new applications. The project has also had a significant impact on university education. Dozens of students have been involved in research, projects and theses in the inter- and transdisciplinary fields of artificial intelligence and mathematics. The launch of the "Mathematics Expert in Data Analytics and Machine Learning" course in English at the Institute is another long-term impact of the project.
The results of the PPKE Faculty of Information Technology and Bionics were presented at prestigious international forums and in peer-reviewed journals, including in the areas of wavelet-based segmentation and multi-discriminator GAN networks. The researchers also focused on proving the origins of neural networks and developing methods to prevent the theft of network weights. They also investigated the vulnerability of neural networks to attack and the reversibility of attacks. Their research in linguistic technology includes automatically generated vector space models of texts and the correction of incorrect, so-called noisy texts that deviate from the colloquial norm. The faculty's achievements also include a patent filed under positive novelty research, the idea for which was motivated by the tragic sinking of the hungarian riverboat Hableány. The aim of the patent is to implement a machine learning system that makes the search and rescue of victims much easier and possible even in difficult circumstances.
The SZTE mainly investigated the interpretability and vulnerability of artificial intelligence algorithms. Both topics relate to the 'black box' problem of artificial intelligence. In natural language processing, interpretable and less resource-intensive report representations and multilingual models have been proposed, and research on the theoretical background of interpretability has been carried out. The researchers identified unprecedented vulnerabilities in the sensitivity of artificial intelligence algorithms to formal verification and have also investigated the limitations of simultaneous attacks on several artificial neural networks. Their results were presented at prestigious international forums. The SZTE also participated in the research of the medical application developed by the Rényi Institue.
Summary of project results:
Science: a total of 79 scientific papers, one patent and one prototype were produced over three years.
Infrastructure: the research infrastructure is in place, typically with high computing capacity servers, which are essential for this research.
Education: during the program, hundreds of university students were introduced to the basics of artificial intelligence by the researchers and educators involved in the consortium.
Innovation: another unexpected result is the emergence of teledermatology. The Covid epidemic made it impossible to take photographs for chronic wound research, but the technology developed – with the help of dermatologists and patients – made it possible to enable remote diagnosis by dermatologists in a very short time. The patient takes a few photos of the area in question, sends them to the doctor, who makes a diagnosis, recommends a therapy and can upload the necessary prescriptions to the cloud – without the patient having to leave the comfort of their home. So far, nearly 20,000 examinations have been carried out in this way.
The next steps
The program has achieved its goal of research into the mathematical foundations of artificial intelligence. This work will be continued on a broader scale by the National Laboratory for Artificial Intelligence (MILAB), also supported by the NKFIH, which will strengthen the synergies between and effectiveness of basic research, applied research and innovation activities.