In 2020, Cisco Systems awarded a two-year advanced research grant to the Icelandic Institute for Intelligent Machines (IIIM) to develop a prototype of a new kind of artificial intelligence (AI).
The new approach, led by Kristinn Thórisson, director of IIIM and a professor at Reykjavik University, differs from existing approaches to AI in several ways. It relies on self-supervised learning, which allows a system to make improvements over time. Learning is based on a form of “reasoning” – where the system autonomously generates hypotheses and tests them.
Also, the new approach does more than simply spot correlations – it recognises causal relations.
Thórisson hopes to develop an AI that can learn from experience in a vast range of situations and transfer its learning smoothly from one context to another. The new AI will even be able to explain why it does what it does.
Weak AI versus strong AI
To understand the significance of the research that Thórisson and his team are doing – with members in Germany, France and Iceland – it is useful to understand the difference between strong AI and weak AI. Strong AI, also known as artificial general intelligence (AGI) or general machine intelligence, refers to a system that can solve problems in multiple domains. Strong AI learns over time through experience.
Today, strong AI exists only in theory. All existing systems are classified as weak AI and can only perform specific tasks in a single domain, such as playing chess or answering questions about a specific product. Weak AI learns through supervised learning, which requires human intervention to prepare training data to help the AI spot the relevant features of the dataset.
Once a weak AI goes through the learning process, it is impossible to predict what the system will do. The datasets used for training are just too large and complicated to be analysed by humans, so the AI often makes decisions that no one fully understands.
Weak AI systems looks for correlations in the data and assume that certain patterns of input will lead to certain patterns in output. According to Thórisson, correlation is not enough; what is required is a system that understands logic and can figure out causation.
Thórisson’s approach, auto-catalytic endogenous reflective architecture (AERA), can change its behaviour on the fly. It takes in new information and “thinks” about what it already knows and what the new data indicates.
Goals are a key element of AERA. Given explicit goals, it compares the goals with actions and outcomes. If a set of actions cause it to reach a given goal, it then tries to determine what actions will lead to a different goal. In this way, the system can be said to think about how it thinks in order to adjust to changing goals.
“Our system reasons using abduction, deduction and induction – and even a little bit of analogy,” said Thórisson. “Abduction is what Sherlock Holmes is so good at. You have a scenario, something has happened. You have a state and you’re trying to infer what happened, how it got to be that way.
“Our approach will yield systems that can come up with new concepts from scratch. These systems will be able to handle unknown sets of variables. If, for example, you have an air traffic control system that detects one more airplane than expected, it doesn’t choke. A key focus of our research is building systems that, through fundamental principles of their operation, can handle the unknown.”
While AERA promises to deliver systems that go way beyond what existing AIs can do, Thórisson’s team is not the only research group exploring strong AI. Another approach that is similar to AERA is called non-axiomatic reasoning system (NARS). This effort has been under development by Pei Wang at Temple University in the US for over 20 years, more recently with the help of a team.
Like AERA, the NARS project hopes to develop strong AI systems that can “think” and follow the same principles as the human mind. Both projects aim to develop systems that can solve problems in a multitude of domains. However, one way that AERA stands out is that it can learn from any new domain – and with a “soft” upper threshold, which means it doesn’t have an upper limit on the number of variables or concepts it works with.
Progress so far on AERA
Thórisson will use the Cisco funds to develop code that will demonstrate his approach more fully, so the two research groups can learn from each other. But this will not be the first demonstration of AERA. About 10 years ago, Thórisson and his team developed a working model that learned how to conduct a mock television interview in real time by observing two people talking about materials recycling.
“It was the first time we really took this seriously and codified our own methodology and followed it to the letter,” said Thórisson. “The system we created surpassed all our highest expectations. It could learn continuously on the fly and could perform unspecified tasks and meet new goals. It could learn by observation from very high-level descriptions of a task.
“This system worked way beyond our wildest dreams. We have spent a lot of time deconstructing what it did to try to condense the principles behind it. Because it’s just so different from the mainstream, it has been quite difficult to explain using only mainstream terminology.”
By developing more code to release to open source – and by running more demonstrations – Thórisson hopes to gain the momentum that will allow him to expand the team and create a community of researchers interested in taking these ideas further.
“It took several years for Wang to create a small team of highly competent people to work on NARS – and that was even after they had a really good code base with an open source version,” said Thórisson. “In the last few years, they have done some really amazing demonstrations, in part with the help of Cisco Systems.
“Cisco is funding my team to do something similar. NARS and AERA are very compatible at the conceptual level and methodologically speaking. There is an opportunity to learn from both systems and bring AI to a new level.”
Thórisson added: “If we can implement just 50% of our ideas, that would be great. That would already be way beyond what current AI does.”