Towards a standard model of mind/common model of cognition

One of the most interesting developments of the last decade within the fields of cognitive AI and cognitive modelling research is represented by the proposal of a Standard Model of Mind (later on called the Common Model of Cognition) by John Laird, Christian Lebiere, and Paul Rosenbloom (Laird, Lebiere, and Rosen- bloom, 2017). The idea of these three researchers was one of abstracting, from some of the most adopted cognitive architectures that they have developed - namely, SOAR, ACT-R, and SIGMA⁴ - a sort of ideal model (a standard one) about the underlying common architectural elements that human-like minds should possess. This abstraction is based on the consensus reached in the community over decades of research and on the convergence reached by these three systems that, despite starting from different assumptions about the architecture of human cognition, have converged towards some interesting commonalities. The areas of consensus reached by such diverse architectures have been grouped into different levels of analysis: (i) Structure and Processing mechanisms, (ii) Memory and Content; (iii) Learning processes, and (iv) Perception and Motor Mechanisms. Concerning the first, an important element of consensus regards the fact that processing in human-like architecture is assumed to be based on a small number of task-independent modules and should support both serial and parallel information processing mechanisms (parallel between modules and serial within them). From an architectural perspective, all the three architectures taken as a source of inspiration for the design of the Common Model converge towards the necessity of distinguishing between a Long-Term Declarative Memory and a Procedural one, as well as the necessity of a working memory module operating as a control interface between the Procedural module and other modules such as the Declarative Memory and the Perception/Motor modules. Concerning the Memory and Content issues, the main point of convergence regards the integration of hybrid symbolic—subsymbolic representations and processing and the inclusion of relevant metadata like frequency, recency, similarity, activation, etc. attached to the processed representations (neural, symbolic, or hybrid). The fact that such integration is necessary is nowadays widely accepted and, indeed, classical symbolic architectures like SOAR are also moving towards hybridization. As for the learning part, the elements of convergence regard the following assumptions: (i) all types of long-term knowledge (i.e., knowledge shared in one of the two long-term memories) should be learnable from a human-like architecture, (ii) learning is seen an incremental processes typically based on some form of a backward flow of information through internal representations

9 Sigma is a novel cognitive architecture that starts with the same basic assumption of SOAR (Rosenbloom, 2013) but that uses probabilistically graphical models, in particular factor graphs, as representational elements in its long-term memory, working memory, and perceptual and motor components. In general, the graphical models can be considered a class of symbolic representations, where the relations between concepts are weighted by their strength, calculated through statistical computations.

of past experiences, and (iii) learning over longer time scales is assumed to arise from the accumulation of learning over short-term experiences. Finally, in regards to the perceptual and motor parts (the least developed, with respect to the others), the main points of consensus are limited to the fact that perceptual and motor modules are assumed to be modality specific (e.g., auditory, visual, etc.) and associated with specific buffers for the access to the working memory. Since all these points of convergence have been reached by starting from completely diverse assumptions, these findings are somewhat surprising and worth investigating. The Standard Model of Mind, however, is currently highly underspecified, as the authors themselves admit. It represents a starting point, a platform for developing high-profile research in both AI and (computational) Cognitive Science. These two disciplines, in fact, can find in projects like this a way to join forces and this can be done in light of a new mutual interest and convenience. While, indeed, AI technology has reached important levels of performance in narrow settings, the missing part concerns exactly the study of how to create artificial companions (embodied and disembodied) that are able to integrate different skills in order to help humans in their everyday activities. Similarly, computational Cognitive Science is interested in individuating how the brain and mind work as integrated systems. This renewed convergence is, in my view, a necessity driven by the fact that modern and future AI and Cognitive Science research will be once again disciplines interested in the same topic: namely, the discovery of the mechanisms that enable multitasking intelligence. In order to advance scientific knowledge in their respective fields, in fact, they need to evolve and become sciences (of the artificial) that study the mysteries of integrated intelligence. The time seems ripe for such a renewed collaboration.