Neurobiological foundations of cognition and progress towards Artificial General Intelligence

The last several decades produced revolutionary advances in two areas: identifying neuronal processes responsible for intelligence and designing machine processes for simulating intelligence. Progress in both areas started in the last century when three foundational ideas were formulated: neuronal assembly as the basic functional unit in the neuronal system, the Turing machine as a universal computer, and the transistor as the basic computing element. Subsequent advances in neuroscience were driven by technologies allowing unprecedented access to processes in the neuronal substrate (e.g. MRI) and by theories hypothesizing general principles of brain operation (the Active Inference principle, the Neuronal Group Selection principle, the Integrated Information principle, other). Progress in machine intelligence involved designing algorithms simulating intelligence functions, in particular, neural network algorithms for machine learning, and designing efficient hardware for implementing the algorithms.

Recent breakthrough accomplishments in language processing (General Language Models, or GLM) resulted from advances in the neural net methods and an almost trillion-fold increase in hardware efficiency which made possible application of these methods in the tasks of realistic complexity. Throughout the history of both areas questions have been raised concerning the relation between them: can machine intelligence profit from using ideas and findings in the study of the brain, can successful methods in machine intelligence shed light on the principles of brain operation and, last but not least, how to structure human-machine interaction to obtain genuine collaboration. Analysis of these issues within frontline theoretical frameworks has revealed divergence: cognitive functions (in particular, understanding) are not amenable to computational simulation (Penrose), and successes in the application of GLM are irrelevant to and shed no light on the origins and mechanisms of natural language (Chomsky). The purpose of this topic is to re-visit these questions, seeking convergence and setting the stage for a possible new synthesis. Expectations of such synthesis derive from the assumption that human intelligence emerged when possibilities for intelligence growth via increasing the number of neurons in the brain and the intensity of their operation were reaching biophysical limits determined by rigid thresholds on the volume of cranial cavity occupied by the brain and the intensity of metabolic energy supply into the brain necessary for its survival and functioning. On that assumption, current trends in machine intelligence relying on packing more processors into chips and intensifying energy supplies are running in the direction opposite to that exploited by evolution.

Accordingly, the benefits of a new synthesis can be dual, providing new insights into neuronal mechanisms of intelligence and new approaches towards expanding the range of intelligence functions delivered on low-to-moderate capacity platforms. Contributions are invited that will address the state-of-the-art in correlating intelligence functions (attention, awareness, perception, learning, understanding, prediction, explanation, consciousness) to neuronal processes, focusing on:

a) Assessing the relative role of these functions in intelligent performance.

b) Mapping them onto principles of machine design. The topic seeks to advance understanding of how evolution obtained indefinite expansion in the variety and complexity of cognitive activities available to the human brain without increasing either the size of the neuronal pool or the energy demands.