In this work, we develop models of neural circuits and plasticity rules that underlie different forms of learning and memory, with a focus on learning processes that involve multiple brain regions. We begin by surveying the literature on synaptic plasticity rules and implementations of learning algorithms in the brain. Each subsequent chapter presents a model of how a specific aspect of learning is implemented biologically, based on experimental evidence and normative considerations.
We first focus on the neural basis of reinforcement learning in the basal ganglia. We show that in order to enable effective learning when control of behavior is distributed across multiple regions (``off-policy reinforcement learning''), classic models of dopamine activity must be adapted to include an additional action-sensitive component. We also show that the known plasticity rules of direct and indirect-pathway striatal projection neurons are inconsistent with existing models of striatal codes for action.
We propose and find experimental support for a new model of striatal activity driven by efferent input. This model is functionally compatible with striatal plasticity rules and enables simultaneous multiplexing of action-selection and learning signals, a necessary ingredient for off-policy reinforcement learning. We next use an off-policy reinforcement learning model to explain a new experimental finding about the conditions under which learned motor skills are consolidated to be driven by the dorsolateral striatum in rats.
We then shift our focus to consider consolidation more broadly, proposing a general model of the advantages of systems in which memories and learned behaviors are consolidated from short-term to long-term learning pathways. In particular, our model proposes that such architectures enable selective filtering of the set of experiences used for learning, which can be essential in noisy environments with many extraneous stimuli.
In the appendices, we explore other factors relevant to learning algorithms, including the interaction between multiple sensory modalities, and the problem of credit assignment in multi-layer neural networks. In summary, this work presents a varied set of models of different forms of learning in the brain, emphasizing the cooperative role of plasticity rules and multi-regional circuit architecture in producing functionally useful synaptic weight updates.
| Identifer | oai:union.ndltd.org:columbia.edu/oai:academiccommons.columbia.edu:10.7916/a2b4-7b42 |
| Date | January 2024 |
| Creators | Lindsey, John William |
| Source Sets | Columbia University |
| Language | English |
| Detected Language | English |
| Type | Theses |
Page generated in 0.0018 seconds