Interview with A. HERZIG and M. SERRURIER, the challenges of artificial intelligence

We publish the third video in a series of interviews presenting research work from our different departments. Andreas HERZIG from the LILaC team and Mathieu SERRURIER from the ADRIA team, researchers in the Artificial Intelligence (AI) department, talk to us about their research in the field of AI, particularly on symbolic reasoning and artificial learning. The challenge? Successfully combining these two aspects of research to converge on hybrid artificial intelligence.

How to define artificial intelligence?

Artificial intelligence is defined as a set of techniques and theories for developing complex computer programs whose goal is to reproduce certain characteristics of human intelligence and to simulate human behavior. Research in artificial intelligence focuses on two distinct aspects that constitute subfields of AI, namely symbolic reasoning and artificial learning. Andreas HERZIG works on the field of symbolic reasoning and knowledge representation. This implies working on both the knowledge acquisition part and developing the agent’s ability to reason from this knowledge. Mathieu SERRURIER focuses on the subfield of artificial learning. He explains that he seeks to build models from data but without any a priori knowledge, which allows these two approaches to be complementary. Building models from data means trying to reproduce desired and/or observed behaviors from a large number of defined values and parameters. The difficulty lies in succeeding in reproducing behaviors without succeeding in understanding the action plan that led the agent to make the right decision.

What are the challenges of artificial intelligence research?

For the past twenty years, artificial intelligence has been a fascinating field that has transformed our societies. Thanks to the discovery of new algorithms and to the mass availability of data, the impact of artificial intelligence is growing, especially since 2007, with the rise of deep learning networks. Deep networks are capable of solving very complex problems, for example in the fields of medicine, translation and autonomous cars. However, even though artificial intelligence technologies can be found in many aspects of everyday life, there are still a number of significant AI challenges. A social aspect of intelligence, fundamental to humans, needs to be built into these technologies. This involves developing the ability of an agent to interact with other agents, to put itself in their shoes, and to understand the goals and beliefs of those agents. Researchers focused on the artificial learning aspect are working on the security of these models, so that they can be integrated into reliable technologies. But the main challenge is to understand how the neural network makes decisions, and this involves developing new specific algorithms.

Today, research is faced with two models, the symbolic reasoning model (precise, powerful, but difficult to apply to real problems) and the artificial learning model, capable of solving very difficult real problems but the reasoning behind it is difficult to interpret. The goal of artificial intelligence research is to combine these methods of artificial learning and symbolic reasoning in order to converge towards a hybrid AI.