Methodology

A “research methodology” is what science calls the methods, approach, tools, and related background assumptions for answering questions about the world. The methodology we choose for studying a phenomenon is extremely important, because without a proper methodology no progress can be made. A good methodology for any subject follows from a good theory of that subject, but since no well-established theory for intelligence exists, we must do the next best thing: We must observe the phenomenon carefully and extract its key requirements. Ideally, we have a complete list of the necessary and sufficient conditions for the phenomenon’s features that we are interested in. But without a proper theory, such a list cannot exist. Again, we do the next best thing: We construct a list of some of the necessary requirements for intelligence (and hope that we get close to it being sufficient). These are some that we consider key:

  • Interacting with a world - i.e. sensing and acting
  • Constructivist cumulative learning
  • Knowledge of causes
  • Autonomy - i.e. self-supervised learning
  • Non-axiomatic Reasoning
  • Domain-independence - i.e. generality
  • Transversal handling of time
  • Transversal resource management
  • Reflection - ability to learn about own cognition

A learning agent whose memory and mind cannot know everything in a changing world must pick and choose what to sense and what to think about. For any goal that such an agent may have, there will always be something that the agent does not know. To find out what the current status of the world is, with respect to its goals, is (a) through interaction with the world. This has also been referred to as “embodiment”. Although this requirement has taken on a life of its own, we cast this simply as a requirement for (b) knowledge of causes. This also means that (c) the agent must learn incrementally, on-the-go (i.e. cumulative learning), and (d) manage its mental resources. Furthermore, this incremental learning and resource management - call it attention - must be independent of the subject of learning (or, as independent as possible), hence the call for (e) domain-independence, or generality. The same can be said about the passage of time: (f) For any piece of knowledge, time must factor in, hence a requirement for (g) transversal handling of time. Now, a complex world can generate a large number of combinatorics; in the physical world we can say that this number is infinitely large. That would call for a lot of memory if we wanted to learn all of it, or even a small part of it! Luckily, a world with regularities, like the physical world, allows us to use logic to compress information about it. But what kind of logic? It cannot be traditional logic, because that kind of logic requires knowledge of its axioms. No, the kind of reasoning needed for an agent like that, in a world like that, is (h) non-axiomatic reasoning. An agent “born” into the physical world will not know everything from the outset, nor will it know all the rules that control how the world behaves. This means that it must be born with the capacity to create new knowledge from scratch. This brings us the rquirement of (i) constructivist learning. Finally, all of this should proceed without the need to “call home”, i.e. (j) autonomosuly. For a system whose cognition improves signficantly with time, a self-monitoring system is necessary, hence the requirement for (k) reflection.

AERA meets all these requirements (and more). Its compositional (transparent) knowledge representation enables it to learn cumulatively and revise its knowledge as it gains more experience. Handling of time is a built-in feature of all knwoledge AERA creates and utilizes. This explicit incorporation of temporal information allows the system to handle dynamic tasks much more easily and naturally than other control systems. AERA uses feedback from the environment, as well as from internal events, at different levels of learning and task performance. For instance, when a model is applied, a number of predictions have taken place, and this information is fed back to the system, which means that the system has internal causal learning. The system also uses feedback about the accuracy of its own predictions, via a meta-control (meta-knowledge) mechanism, to improve its knowledge.

Related Material
2020 Constructivist AI - Lecture Notes at Reykjavik U. HTML
2017 Architectures for Generality & Autonomy: Which Methodology? PDF
2014 What Should AGI Learn from AI & CogSci? Video

Causation

Intelligence is about getting things done! So, it calls for information about what leads to what, which is what has been called causation. Without causal knowledge representation explainability, knowledge transfer, counterfactual reasoning, and in general, understanding becomes impossible.

In AERA system, causal models are fundamental components of knowledge representation. AERA learns causal models in real-time through performing experiments. The learned models are used for making predictions (forward chaining) and goal achievement (backward chaining). In fact, AERA creates causal graphs of knowledge that demonstrate how to perform a task, through different potential paths of achieving active goals. There also exist other higher-level causal models that explain failures, if particular models fail to predict, by investigating the potential causes of the failures.

Constructivism

Constructivism, as a psychology theory, states that “the mind of humans, actively develops its cognitive capabilities via a self-directed constructive process” (Piaget 1951). The theory, which was then brought to AI, has a particular focus on autonomy of agents. [Link to Kris’s paper]:

Non-axiomatism

AERA assumes that task-environments are non-axiomatic (like the physical world). Therefore, its knowledge representation and reasoning are non-axiomatic as well. The inferences made by a non-axiomatic reasoner never go beyond what the experience supports. So, the causal models learned by AERA are only based on what the system itself experiences during its lifetime. AERA’s learning occurs in a self-supervised manner, and since the system has no strong attachment to its prior knowledge, it has a high potential of dealing with novel situations.