Artificial General Intelligence (AGI) requires an infrastructure, for algebraically expressing composable structures and operations, built on powerful abstractions. In a well-trained system built on these principles, learning to make appropriate algebraic substitutions would be sufficient to make intelligent responses. The power of Category Theory (CT) to dynamically compose appropriate granular units of computations and Homotopy Type Theory (HoTT) to help with semantics and validate the compositions, is necessary for this. This paper describes the roles CT and HoTT could play in providing a mathematical foundation for implementing a framework suggested by the Subjective Emergence of Objective Minds (SEOM) model developed earlier by the author. The paper concludes by observing that insights derived here, regarding Deep Neural Networks (DNN) and other related approaches to artificial intelligence, are necessary to put us on the correct path to AGI.

An Artificial General Intelligence (AGI) system must intelligently deal with the semantics of the situations it encounters. The flexibility required for intelligent responses is at odds with the fixed parameter spaces of statistical methods or the limitations of the rule-based syntax of the computing paradigms on which systems can now be built. Overcoming these limitations requires philosophical insights from the work of Kant, Wittgenstein, Heidegger, and Gadamer, covering ontology, epistemology, hermeneutics, axiology, and teleology. Specifically, additional teleological computations (to determine the purpose served, independently from the computations responsible for responding appropriately in the world), are necessary to overcome these shortcomings.These insights suggest using a new abstraction to satisfy the necessity for two computational processes to execute in parallel on their respective, related but different representations. The processes must consist of algorithms specifiable a priori, to apply in all possible situations. In this paper, we address some specific difficulties with concepts and show how these difficulties may be overcome using the new abstraction of higher-level perceptions proposed in the Subjective Emergence of Objective Minds (SEOM) model developed earlier.In this paper, we will focus on the representations and teleological computations necessary to give purpose to the actions of autonomous intelligent agents.

The ‘brain-mind-intelligence’ structure may be considered analogous to a ‘computer-operating system (OS)-application’ construct. Software development would be far more difficult without the benefit of abstractions like file and memory handles, and graphical user interfaces, implemented in operating systems. The easy access to application-level semantics, provided by the infrastructure of the OS, simplifies application development.  Abstractions can play a similar role in producing intelligent systems. The power of abstractions may be determined by their ability to unify the treatment of several types of higher concepts. Unifying various concepts in the domain of intelligence, such as problems, solutions, objects, and emotions, through abstractions, will be of immense value. This paper shows how ‘higher-level perceptions’ could be used to serve this objective and construct an OS-like infrastructure. With the infrastructure in place, ordinary software objects developed and deployed using standard methods, become accessible using the semantics provided by the infrastructure. Now, after brief training exercises, the objects would be available for intelligent use. The focus, therefore, is on the design of the infrastructure that could play a mind-like role in humans, using some ideas from philosophers like Kant, Wittgenstein, Heidegger, and Popper. The design uses a new, non-symbolic, non-connectionist, transmutable method of representation that tackles the so-called ‘frame’ and ‘grounding’ problems of artificial intelligence. The infrastructure enables the conversion of sensory inputs to transmutable representations (amenable to semantic modifications), to flexibly act on software objects developed externally. Furthermore, the infrastructure serves to ground and functionally manage the transmutable, connectionist, and symbolic systems. Thus, the new model plays a critical role in building intelligent systems that overcome major problems in learning and action selection, including brittleness associated with the limited semantics of purely symbolic and connectionist designs. The new model is then evaluated in terms of features, against criteria set forth independently by Newell, Sun, Vernon, and others. In addition, a new measure of intelligence introduced in this paper is used to quantitatively compare the proposed model with some others developed over the years.