Towards Cognitively Plausible Associative Memories Capable of Learning Complex Tasks

Abstract

This thesis explores the development of artificial neural networks (ANNs) grounded in cognitive principles. Specifically, it focuses on Associative Memories (AMs) and how to improve their learning properties while retaining cognitive plausibility. To do so, gradual architectural and functioning changes are explored, ultimately leading to a framework for multilayered and multinetwork AMs capable of learning complex tasks while sharing the same underlying principle. Chapter 1 investigates the variability and consistency within an unsupervised AM in categorizing similar and distinct stimuli, demonstrating that while individual representations vary, the learning outcomes are consistent across networks. Chapter 2 combines unsupervised AMs to create a multilayered architecture capable of learning and gradually distinguishing highly correlated inputs. Chapter 3 integrates this unsupervised architecture into a Bidirectional Associative Memory (BAM) to form a multilayered BAM capable of learning non-linear tasks relying solely on local associative learning processes. Chapter 4 presents a novel binary transmission function for AMs that improves biological plausibility while maintaining high learning performance. It also investigates the new properties gained by binary encoding, such as information flow control and memory replay, which facilitate dynamic interactions between multiple networks. The thesis concludes by discussing the implications of these findings for artificial intelligence and cognitive modelling and suggests future research directions for developing more robust and adaptable neural network architectures. It emphasizes the growing importance of multidisciplinary collaboration across various perspectives and fields in unravelling the complexities of brain function.