Researchers claim that the model, designed using analysis of existing systems and a wide range of research in this area, will form the basis of future artificial intelligence (AI), which will be “error-proof” and capable of abstract thinking. The research paper was published in the Mind and Matter (Journal).
AI systems help humans facilitate and automate decision-making in a multitude of areas, from housekeeping to warfare. In the future, AI will make increasingly more decisions without human intervention, the UTMN scientists said. This creates the need for approaches to AI architecture that are guaranteed to avoid unacceptable risks.
Robots and AI systems used today, for example to drive cars, predict consumer behaviour, and even adjudicate, derive their answers statistically, based on an array of analysed examples. These systems can be very effective in certain computational tasks, but they are still far from “general”, or “strong”, artificial intelligence. Therefore, the scientists stated that it is not uncommon for AI to make incorrect and unethical decisions.
As long as AI is guided by a given set of samples with no ability to re-evaluate or expand it in virtually real time, there will be a risk that the automatic decision will turn out to be wrong, the UTMN scientists said.
To mitigate such risks, the university’s researchers have proposed an approach to developing AI that will make it more intelligent and “conscious”. According to the authors, their AI architecture model combines the results of a wide range of physical, mathematical, cognitive, philosophical, and other research in the field.
“We have derived a key attribute of having a ‘general AI’ system – it must be able to understand and apply different models of its environment, simply put, different ‘worldviews’ or ‘theories’. Driven by ‘theories’ as particular paradigms for evaluating data, AI will not only be able to ‘look at a situation differently’ but also derive new principles, laws that are relevant to the field”, said Louis Vervoort, a researcher at the UTMN School of Advanced Studies.
Abstract thinking for both humans and machines means using consistent models and conceptual frameworks that allow you to study the world from different angles, expanding your interaction with it, the scientists explained.
“The ethical issue of AI is not whether machines lack a common ‘human’ understanding of the world that includes unambiguous moral attitudes and the idea of moral responsibility. High-level cognition, consciousness and ethical behaviour stem from the same source – the ability to master broad conceptual frameworks, not just specific tasks. We believe that this process can also be reproduced in a machine”, Louis Vervoort explained.
The proposed approach has some similarities with David Rosenthal’s high-order-thought (HOT) theory of consciousness, as well as with Bernard Baars’ Global Workspace Theory (GWT). According to UTMN scientists, these theoretical models of consciousness and abstract cognition were supported by recent experimental research in cognitive neurobiology at the Massachusetts Institute of Technology (Cambridge, Massachusetts, USA).
Some of the most advanced AI systems, mimicking to some degree abstract thinking, were produced by physicists at the Swiss Federal Institute of Technology (Zürich, Switzerland) and the Massachusetts Institute of Technology. These studies demonstrated AI's capacity to independently formulate and combine physical laws – the first step towards theories.
“So far, we have suggested only a general strategy. But our analysis makes it clear how advanced modern AI is in imitating or implementing abstract thinking. If you look closely at the details, there is still a long way to go before ‘theoretical prowess’ is achieved. However, we believe that future AI should become ‘theory-driven’ and not ‘data-driven’ as it is now”, Louis Vervoort said.
The researchers believe that the proposed approach would form the basis for the development of the next generation of AI. In the future, the research team plans to continue working closely with experts in artificial intelligence and machine learning to develop specific computational approaches based on the new model.