Towards a Predictive Processing Implementation of the Common Model of Cognition

Modern machine learning techniques based on artificial neural networks (ANNs) are implemented through algebraic manipulations of vectors, matrices, and tensors in high-dimensional spaces. While ANNs have an impressive ability to process data to find patterns, they do not typically model high-level cognition. Furthermore, ANNs are usually models of only a single task. Otherwise, when an ANN is trained to learn a series of tasks, catastrophic interference occurs, with each new task causing the ANN to forget all previously learned tasks [8, 21, 22]. On the other hand, symbolic cognitive architectures, such as the widely used ACT-R [1, 31], can capture the complexities of high-level cognition but scale poorly to the naturalistic, non-symbolic data of sensory perception, e.g., images, or to big data sets necessary for modelling learning over a lifetime, e.g., corpora with hundreds of millions of words.