Aiming at reasoning AI
Researchers at Luleå University of Technology try to create systems that match the abilities of animals and humans. Their goal is to contribute to the solution of Artificial General Intelligence. – For humankind, it is the most interesting problem in the world to solve, says Dmitri Rachkovskij, visiting Professor at Luleå University of Technology.
The challenge is to implement human reasoning into the neural network based Artificial Intelligence. This is done by using concepts from neural networks and combining them with symbolic reasoning: fundamental basic research that belongs to the emergent field known by the name Hyperdimensional Computing or Vector Symbolic Architectures (HDC/VSA).
– We unify different branches of AI with our approach. By doing that, we can create more general systems with more capabilities, says Evgeny Osipov, Professor of Dependable Communication and Computation Systems.
Difficult translations
Imagine a language translation service like Google Translate. There, words are encoded by the AI algorithm such that a good part of their meaning is preserved – and that makes such services so popular today. However, from time to time they produce wrong translations with an incorrect meaning. This is because the current AI engine behind such services does not apply the world knowledge to understand and reason about the situations described in the given text.
– This is an example of limitations in modern AI systems; they cannot reason about what they deal with, says Evgeny Osipov.
Autonomous systems
The field of HDC/VSA unifies two main approaches for creation of Artificial Intelligence, the symbolic approach and the neural network or connectionist approach. The symbolic approach can be described as top-down, from function to structure. For example, if we identify a function of natural intelligence that we want to mimic, we implement it with some computer algorithm operating with symbols and numbers. The other approach has a bottom-up perspective, from structure to function. In this approach, we build models of brain neural networks, often in terms of connected artificial neurons, and defines its learning rule, usually modifying the strengths of connections. Basically, HDC/VSA combines the advantages of bottom-up and top-down approaches to AI, while trying to avoid their drawbacks.
– As a more applied and basic example of our research, we develop methods for combining object representations from various modalities. Simply speaking, a visual representation of a cat is combined with its acoustical and tactile representations. HDC/VSA make it possible by using high-dimensional distributed vector representations, and without any training, says Dmitri Rachkovskij.
– Hopefully, our research can contribute to future autonomous systems that can work without intervention from humans, by combining neural network representations with the ability to reason, concludes he.
Dmitri Rachkovskij used to work at the International Research and Training Center for Information Technologies and Systems, a research organization of the National Academy of Sciences of Ukraine in Kyiv. During many years, he wrote joint publications with researchers at Luleå University of Technology. After the war began in Ukraine, Dmitri Rachkovskij came to Luleå University of Technology via funding from the Swedish Foundation for Strategic Research program, the Swedish Research Council and Luleå University of Technology.
Read more about HDC/VSA and about how we can reach resoning AI:
Updated: