Kurzweil 2012 Awards Interview

Kristinn R. Thórisson, IIIM Managing Director, and his collaborators Helgi Páll Helgason and Eric Nivel, received the 2012 Kurzweil Award for Best AGI Idea at this year’s AGI conference at Oxford University in Cambridge. We interviewed two of the paper’s authors, Helgi Páll Helgason and Kristinn R. Thórisson, regarding the award and what the next steps in their research will entail.

On Attention Mechanisms for AGI Architectures

Helgi Pall photograph
Helgi Páll Helgason.

The paper, titled “On Attention Mechanisms for AGI Architectures: A Design Proposal” (download from Mindmakers.org), presents arguments for super-intelligent artificial agents needing what we generally think of “attention”, and presents a blueprint for how to achieve the design and implementation of such attention mechanisms.

What significance does this award have for your future research?
Helgi: This is an invaluable validation of our work up to this point and strong motivation to further pursue research on control mechanisms and resource allocation for AGI’s. Our chances of getting our voices heard in the scientific community, finding funding for future research as well as new collaborators are also positively impacted by the award.

Kristinn R. Thorisson photograph
Kristinn R. Thórisson.

Kristinn: More than any sub-topic of AI I have worked on in the past 20 years, with the possible exception of my Ph.D. work in the early ’90s, this work has seemed completely isolated from anything and everything that everyone else is working on within the field. This could signal that either we are simply nuts – working on something of no importance and no significance – or, that we have identified an important research topic that everyone else has missed. This award puts this work squarely in the second camp. On top of that it is always nice for one’s work to be recognized by one’s peers.

What are the next steps now that you have designed an attention mechanism?
Kristinn: Over the past four years we have developed a unique computational foundation for AGI that I hope will not only get a one-off recognition at a single conference but have a transforming effect of the field as a whole. Our AERA cognitive architecture is the first working implementation of an auto-catalytic, constructivist system resulting from constructivist methodology, which is a radical departure from the methodologies everyone else is using. Getting the methodology right is a necessary – but of course not sufficient – prerequisite for achieving some results in any field, and AI is no exception. We anticipate a paradigm shift over the next 10 to 20 years, away from constructionist AI to constructivist AI, which will eventually lead to the achievement of artificial general intelligence, and show that strong AI is not only a theoretical possibility but a practically achievable one as well.

Helgi: My short term focus will be on completing evaluation of the mechanism and completing my dissertation. Beyond that, experiments on more tasks and domains will undoubtedly lead to improvements and new insights.

Do you have any comments on the importance of this article within the field of AGI?
Helgi: Establishing attention as a critical function of AGI systems is important, as any practical applications in the real world will require this capability. It has been all to common to see entire architectures being designed and implemented without much regard for the constraints imposed by the complexity of real world environments and realtime processing. Considering that the our brains require the enormous information reduction carried out by human attention – and that modern computer hardware is probably at least a decade a way from matching our best estimates of the computational power of the brain – the necessity of attention for AGI is not hard to argue for. As discussed in our paper, attention is a notoriously difficult capability to retrofit existing architectures with due to its pervasive nature and the fine temporal granularity required. Hopefully our work will have some influence on the design of current and future AGI architectures in such a way that these issues are given proper consideration from initial stages.

Kristinn: At CADIA we are slowly but surely filling in some of the major gaps that the field of AI has left wide open for the past 50 decades, and attention is a good example of that. In our approach AGI systems must be built on the assumptions that such systems “create their own knowledge”, that is, that their understanding of the world is actively constructed through their own interaction with the world. As Pei Wang (collaborator and Affiliate Researcher at IIIM) has argued for over a decade, AGI systems working in an environment of sufficient complexity will always be mentally underpowered – like humans, they will always have insufficient knowledge and resources with respect to their goals in the world. To take an example, you can never be certain – absolutely 100% certain – that you will not be killed in an accident tomorrow while going about your daily errands. Yet most people would give quite a lot to have that certainty. When you take the assumption of insufficient resources and knowledge into account you soon come to realize why animals have what we call “attention” – the ability to sensibly steer the use of their mental resources to things and areas of importance.

What would be a typical real-world application for your ideas; how will this potentially result in more capable AGI systems?
Helgi: This work can only be practically realized as part of an AGI system. This has already happened with the AERA architecture developed in the HUMANOBS project. The question then boils down to typical real-world applications for AGI systems. The Wozniak test describes a task for an embodied robot, where the robot should be able to enter any randomly chosen household and make a cup of coffee. While the task sounds simple, at the present stage of AI research such a robot would be practically impossible to build due to the rich variations that exist between households and the fact that the goal is stated at a high level of abstraction. Any system capable of solving such problems set in real-world environments requires attention functionality as the total information available from the environment is vastly beyond what could be processed at depth in real-time. So attention does not only result in more capable AGI systems, it is a requirement for AGI systems that are capable at all of solving problems of real-world complexity. If the coffee making task does not sound like a big deal, consider that the capability to solve this task will also imply that the capability to solve a wide range of considerably more important tasks will be well within reach, from such as building autonomous robots capable of saving trapped humans from any burning building to robots that can autonomously explore outer space.