Connectionism and Learning: Neural Networks Explained

Introduction to Connectionism

This paper examines connectionism and its potential to model various learning processes in the brain through a multidisciplinary approach that combines many different theoretical frameworks — an approach that has recently received a significant boost from advances in technology. The basic principles of the connectionism model involve a sense of biological realism built upon interconnected networks that form a more complex whole, potentially explaining the processes within the human brain. These same principles may also serve as a foundation for developing non-human networks, such as artificial intelligence. Although it is not entirely clear how this research might be relevant to specific career goals at present, the field is developing so rapidly that knowledge of this subject could become highly relevant within the next five years.

Citation: Mayor, J., Gomez, P., Chang, F., & Lupyan, G. (2014). Connectionism coming of age: legacy and future challenges. Frontiers in Psychology. doi:10.3389/fpsyg.2014.00187

The History and Future of Connectionism (Mayor et al., 2014)

This article was chosen because it provides a rich and detailed history of how the connectionist model developed, as well as discussion of where the model could advance in the future. In 1986, Rumelhart and McClelland introduced the Parallel Distributed Processing (PDP) framework to the cognitive science community — a framework that sought to construct, at the algorithmic level, models of cognition compatible with their implementation in a biological substrate (Mayor, Gomez, Chang, & Lupyan, 2014). After reviewing the obstacles the theory has encountered, the article identifies its key challenge — learning abstract structural representations — and explains how many gaps must be filled before this model can fully account for complex intelligence.

Citation: Dalege, J., van den Berg, H., Borsboom, D., Conner, M., & van der Mas, H. (2016). Toward a formalized account of attitudes: The Causal Attitude Network (CAN) Model. Psychological Review, 2–21.

This article evaluates the ability of the CAN model to explain the reactions people have to events, as well as the interactions among those reactions. For example, when a person recoils at the sight of a snake, they are not engaging in a rational internal assessment of the threat; rather, they react based on their "attitude" toward snakes. Understanding how these attitudes are structurally represented in the brain's neural networks can begin to illuminate how such reactions manifest from a connectionist perspective and potentially open avenues for further study.

Attitudes as Neural Networks: The CAN Model (Dalege et al., 2016)

Citation: Plaut, D., & Vande Velde, A. (2017). Statistical learning of parts and wholes: A neural network approach. Journal of Experimental Psychology, 318–336.

This article examines the phenomenon of statistical learning and considers how insights from statistical learning research might help identify the properties neural networks must possess in order to learn various types of material. For example, learners are sensitive to different "chunkings" of information drawn from multiple modalities, including auditory and visual sources. The researchers construct a Bayesian statistical model to test whether such insights can help resolve some of the persistent obstacles in understanding neural network learning.

Citation: Nelson, R. (2013). Expanding the role of connectionism in SLA theory. Language Learning, 1–33.