Abstract
Animals learn about their surroundings and associate actions with their outcomes. Associative learning, facilitated by synaptic plasticity, enables the learning of individual events as well as the association of two events. In this dissertation, we explore how events are associated together when they are separated by a long delay. We propose reactivation as a possible mechanism through which two temporally separated events could be linked together. We also examine how credit is distributed when multiple actions could have contributed to the outcome.
In Chapter 2, we study how reactivation can help associate a conditioned stimulus (CS) with an outcome (unconditioned stimulus, US, such as malaise), which occurs after a long delay. Timescales of synaptic plasticity cannot explain such an association across many hours. Most plasticity rules range from a few milliseconds to a few seconds in their windows of associability, over which co-activity between neurons would result in changes in the synaptic weights between them. We hypothesize, based on related experimental evidence, that reactivation of memories of the CS during the US could make this association possible.
We study the case of serial overshadowing, where another novel stimulus, the interfering stimulus (IS), is introduced after the CS, before the malaise. The credit for sickness is distributed across the two stimuli, and depending on the timing of the IS and the US, different levels of conditioning towards the CS, which is overshadowed by the IS, are observed. We explore how the competition between the memories of the CS and the IS could help explain how the relative credit is assigned in the case of overshadowing.
In Chapter 3, we investigate how multistability can emerge in random networks even when there is no self-excitation to make the units individually stable. We see that the response function shapes the regimes of network dynamics. Through analysis and simulations, we show that multistability can arise from the recurrent interactions in the network. We also observe that finite-size networks exhibit multistability in regimes where infinite-size networks do not.
In Chapter 4, we discuss concluding remarks for Chapters 2 and 3 and discuss possible future directions. We outline three possible investigations that would help understand and characterize long-delay credit assignment, as well as shed light on the dynamics of memory interference.