Research
» Novamente Cognition Engine
The Novamente Cognition Engine is a software system designed for large-scale implementation on a distributed network, and founded on a unique design for machine cognition. The overall architecture of the system is based on a unique theory of intelligence founded in complex systems theory and cognitive science; but the specific mechanisms used within the system are drawn from cutting-edge computer science, drawing on a number of recent advances in machine cognition theory and practice.
Currently NCE is being used to control a humanoid agent in the AGISim 3D virtual-reality simulation world. In this context, the NCE has learned to carry out a variety of tasks (such as word-object association, “fetch”, “tag” and object-finding) identified by Piagetan developmental psychology as typifying the "infantile" stage of development. Informally, this means that the system now operates as a simple “artificial baby.”
According to the cognitive theory underlying the NCE, robust general intelligence can only be achieved by a software system with a powerful reflective capability – the capability to creatively analyze and improve its own performance, with a view toward more optimally achieving its goals. This sort of reflectiveness requires that a software system be able to develop what cognitive scientists call a "self-model": an adaptively updated model of the system itself based on its impact on its environment and its interaction with other intelligent agents. But the construction of an adequate self-model requires more powerful and scalable cognitive algorithms than has been present in prior machine cognition or machine learning software systems. Put simply, Novamente is the first software system to possess a sufficiently sophisticated combination of learning and reasoning algorithms to be able to understand itself well enough to progressively improve itself.
Cognitive Architecture
Among the key cognitive mechanisms of the system are a probabilistic reasoning engine based on a novel variant of probabilistic logic called Probabilistic Logic Networks; an evolutionary learning engine that is based on a synthesis of probabilistic modeling and evolutionary programming called MOSES, pioneered in Moshe Looks’ 2006 PhD work at Washington University; and an artificial economics based system for attention allocation and credit assignment.
These mechanisms are integrated using a knowledge representation in which declarative knowledge is represented using weighted, labeled hypergraphs; and procedural knowledge is represented using programs in a customized functional programming language called Combo; and mechanisms are in place for freely converting between declarative and procedural knowledge.
The overall architecture consists of a collection of specialized units, each of which uses these same representations and cognitive mechanisms to achieve a particular aspect of intelligence, such as perception, language learning, abstract cognition, action selection, procedure learning, etc. The system is designed so that the coordinated dynamics of the mechanisms and components will be able to give rise to the emergent structures and dynamics associated with intelligence, including a sophisticated self-model and an appropriately adaptive “moving focus of attention.”
For more, see our Videos, Papers and AGIRI.