Icelandic researcher advocates for an overhaul of artificial intelligence

What are the key attributes of future AI? 

According to Thórisson and Minsky, autonomous general learning, or general self-supervised learning, involves creating knowledge structures about unfamiliar phenomena or real-world objects without assistance. An AI needs to represent cause and effect and use that as a key component in its reasoning processes. When confronted with a new phenomenon, the AI should be able to develop a hypothesis about causal relationships. 

“The most important ingredients for future general machine intelligence are the ability to handle novelty, the ability to manage experience autonomously, and the ability to represent causal-effect relationships”

An AI must be capable of learning incrementally, modifying its existing knowledge based on new information. Cumulative learning involves reasoning-based acquisition of increasingly useful information about how things work. Models should be improved when new evidence becomes available. This requires hypothesis generation – a topic for future AI research. 

What is already known, though, is that hypotheses should be formed through a reasoning process that includes deduction, abduction, induction and analogy. And an AI should keep track of arguments for and against a hypothesis. Another important requirement of a future AI is that it should model what is not known at the time of planning. It should then be able to bring useful knowledge to a task at any point in time.  

“The most important ingredients for future general machine intelligence are the ability to handle novelty, the ability to manage experience autonomously, and the ability to represent causal-effect relationships,” says Thórisson. “A constructivist approach to AI already provides a useful starting point for addressing the first two points – handling novelty and managing experience autonomously. However, we still have a long way to go before systems can model causality autonomously, in an effective and efficient manner.” 

How do we get there from here? 

The current generation of artificial intelligence systems uses a constructionist approach, which Thórisson says has resulted in a diverse set of isolated solutions to relatively small problems.  

“AI systems require significantly more complex integration than has been attempted to date, especially when transversal functions are involved, such as attention and learning,” he says. “The only way to address the challenge is to replace top-down architectural development methodologies with self-organising architectures that rely on self-generated code. We call this ‘constructivist AI’.”

Both Thórisson and Minsky are working on algorithms based on these principles. Thórisson demonstrated an approach to constructivist AI with a system, known as AERA, that autonomously learned how to participate in spoken multimodal interviews by observing humans participate in a TV-style interview. The system autonomously expands its capabilities through self-reconfiguration.  

AERA, which has been under development for 15 years, learns highly complex tasks incrementally. Starting with only two pages of seed code for bootstrapping and running on a regular desktop computer, the AERA agent created semantically meaningful actions, grammatically correct utterances, real-time coordination and turn taking – without learning anything beforehand. It did this after only 20 hours of observation. The knowledge produced consisted of over 100 pages of executable code that the system wrote on its own to allow it to take either role of interviewer or interviewee.  

Focusing on industrial automation, Minsky is taking a very similar approach to intelligent systems at Leila AI. Its neuro-symbolic technology has resulted in a new approach to industrial automation that can track the activities of people and machines on a factory floor and produce actionable information about their operations.  

According to Minsky and Thórisson, the current focus on deep neural networks is hampering progress in the field. “Being exclusively dependent on statistical representations – even when trained on data that includes causal information – deep neural networks cannot reliably separate spurious correlation from causally-dependent correlation,” says Thórisson. “As a result, they cannot tell you when they are making things up. Such systems cannot be fully trusted.”