Adaptron Cognitive Architecture

The Adaptron Cognitive Architecture and Binary Neurons as a General–Purpose Representation

 

Adaptron is a cognitive architecture that is designed to control the intelligent behaviour of robots. It uses compositional hierarchies of binary neurons (binons) as its representational system. Binons are general-purpose, relational and functional nodes for representing knowledge, concepts and abilities. Adaptron satisfies many of the important requirements for artificial general intelligence. These requirements include purposeful, grounded, autonomous, general-purpose, scalable and reliable. Experiments have shown that this architecture is effective for the recognition of handwritten digits and Morse code as well as for the control of a simulated robot in a maze environment.

Adaptron interacts with its environment via senses and action devices. As it learns, it builds up integrated perception–action hierarchies of binons to represent its experiences. This mental model of its world is then used for thinking and rehearsing action outcomes. It is also used to control mental operations such as paying attention, selecting actions and reasoning. Binons are used to perform all of these operations.

A binon is a simple deterministic artificial neural node that represents a relationship. It contains an integer value used to help represent things and their relationships. It has links to two lower nodes and it is reused by zero or more upper nodes. Binons are general-purpose components that interact with each other like objects in object-oriented software. There are currently four types of binons.

  • Property binons are used to represent sense independent property values such as position, intensity, time, repeat counts and the properties derived from them. 
  • Entity binons represent types of things such as objects, events, actions and concepts. Entity binons are grounded on property binons.
  • Control binons are used to learn, manage and repeat behavioural and mental processes.

Binons can also be subdivided based on their purpose.

  • Perception binons are used in recognition and prediction.
  • Action and expectation binons are used for behavioural control. They are equivalent to command neurons in neuroscience, production rules in cognitive science, or the forward and inverse models in motor control.
  • There are mental operation binons to focus attention, perform reasoning and initiate actions.

Adaptron starts with no knowledge or abilities. New binons are continuously created and integrated with existing ones to represent everything it learns. The resulting networks are deep overlapping compositional binary hierarchies. Learning takes place in five stages: reflexes, babbling, reuse, practice and automaticity. Novel experiences result from reflexes and babbling and they become familiar through reuse. They become more reliable when practiced. They then can be performed as automatic habits. This is consistent with the dual process theory of cognition.

If you are interested in reviewing the book, I can provide you with a draft version, one chapter at a time. The book currently has 10 chapters. The first chapter is the introduction and executive summary. It is available here as a PDF file: Adaptron and Binons - Introduction, V1.1.pdf
Chapter 2 and 3 are also now available for review.

If you provide useful feedback on a chapter, I will send you the next chapter when it is ready.

Please respond by using  Contact us  below.