Emotion and Learning:
A Computational Model of the Amygdala

Morén (2002). Emotion and Learning - A Computational Model of the Amygdala. Lund University Cognitive Studies, 93. Lund: LUCS.

PhD Thesis

ISBN 91-628-5212-4
ISRN LUFDA/HFKO-1010
ISSN 1101-8453

Download PDF.

Abstract

The amygdala is a small subcortical structure that has long been implicated in the conditioning of fear and other emotions. It is heavily interconnected to a number of both cortical and subcortical structures and is thus well placed to integrate  sensory inputs from multiple areas to produce emotional reactions directly as well as influence learning and attention systems. Data suggests that the amygdala works in close cooperation with the orbitofrontal cortex; the amygdala learns emotional reactions to stimuli, while the orbitofrontal cortex learns to inhibit the reactions from the amygdala in a context-sensitive manner. The hippocampus is encoding the contextual representations that are used by the orbitofrontal cortex. Being responsible for the conditioning of emotional reactions, the amygdala forms a part of a conceptual system integrating emotions, motivation and actions. The thesis briefly discusses this system, and also reviews the neurophysiological and neuroanatomical features of the amygdala, the orbitofrontal cortex and related areas.  

As a learning system, data suggests the amygdala is working as a classical conditioning system. This system is used both to elicit autonomous reactions to emotional stimuli directly through the central amygdala, and as an evaluative part of an instrumental conditioning system through the basolateral amygdala. Through this mechanism, the structure is also involved in selective memory consolidation and in selective priming of stimuli in the sensory cortices. Classical and instrumental conditioning is discussed, and a number of computational models of classical conditioning are presented and compared. 

The second half of the thesis presents a computational model of the amygdala and orbitofrontal cortex. The model has a very simple design for each component; much of the abilities of the model instead comes from the neuroanatomically guided interconnections between these components. The model is tested with only the amygdala and orbitofrontal cortex, and then extended with a simple hippocampal model able to generate contextual signals from an externally imposed attentional sequence. The model is compared to the previously tested conditioning models and its benefits and drawbacks - especially its current inability to handle time-dependent effects - are discussed.