Are Associations All You Need? Overcoming the Problem of One-to-Many Association and Nonlinear Relationships with Contiguity in a Bidirectional Associative Memory Framework

Abstract

This thesis investigated how mechanisms related to contiguity could help explain the link between biology and cognition through an information signal processing model. It focused on using recurrent neural associative memories, specifically, a biological and cognitively plausible Bidirectional Associative Memory framework. The objective was to overcome its current limitations in learning one-to-many associations and nonlinear relationships. It proposed two major changes to achieve this: the use of contextual information and a new architecture inspired by the laminar organization of the brain. In Chapter 1, contextual units were added to the input layer to change its overall representation. Results showed that by adding context, it was possible to manipulate the inter- and intra-group correlation during an association task. Of importance, it enabled the ability to learn one-to-many associations within the framework. Chapter 2 introduced the multi-feature extraction bidirectional associative memory, a multilayered model designed to investigate if nonlinear tasks, such as the N-bit, double moon, and its variants, could be solved within the framework. Results showcased how this recurrent multi-structure could generate various representations from a single signal. These were used to solve the nonlinear tasks. Results also showed how a transition from linear to nonlinear answers could be achieved within the proposed architecture. Finally, Chapter 3 employed two of these models and used contextual information to have them interact to solve complex cognitive tasks, such as the N-bit and dimension change card sort. Results showed that by using interacting associations, the joint model could learn rule-like behaviors and solve both tasks more efficiently and generally. This thesis showed how the simple mechanism of contiguity could account for complex cognitive behavior. It showcases how associative memory mechanisms can emulate rule-like behavior and surpass the traditional rote association strategy. This avenue holds great promise for enhancing the generalization ability of artificial neural networks. Of interest is that it also showed abundant evidence for new accountability of the associative memory mechanism. This paves the way for a deeper understanding of cognition and how information signal processing, through mechanisms related to contiguity, can amount to a human’s vast repertoire of cognitive abilities.