2024 Theses Doctoral
Synaptic plasticity and memory addressing in biological and artificial neural networks
Biological brains are composed of neurons, interconnected by synapses to create large complex networks. Learning and memory occur, in large part, due to synaptic plasticity -- modifications in the efficacy of information transmission through these synaptic connections. Artificial neural networks model these with neural "units" which communicate through synaptic weights. Models of learning and memory propose synaptic plasticity rules that describe and predict the weight modifications. An equally important but under-evaluated question is the selection of \textit{which} synapses should be updated in response to a memory event. In this work, we attempt to separate the questions of synaptic plasticity from that of memory addressing.
Chapter 1 provides an overview of the problem of memory addressing and a summary of the solutions that have been considered in computational neuroscience and artificial intelligence, as well as those that may exist in biology. Chapter 2 presents in detail a solution to memory addressing and synaptic plasticity in the context of familiarity detection, suggesting strong feedforward weights and anti-Hebbian plasticity as the respective mechanisms. Chapter 3 proposes a model of recall, with storage performed by addressing through local third factors and neo-Hebbian plasticity, and retrieval by content-based addressing. In Chapter 4, we consider the problem of concurrent memory consolidation and memorization. Both storage and retrieval are performed by content-based addressing, but the plasticity rule itself is implemented by gradient descent, modulated according to whether an item should be stored in a distributed manner or memorized verbatim. However, the classical method for computing gradients in recurrent neural networks, backpropagation through time, is generally considered unbiological. In Chapter 5 we suggest a more realistic implementation through an approximation of recurrent backpropagation.
Taken together, these results propose a number of potential mechanisms for memory storage and retrieval, each of which separates the mechanism of synaptic updating -- plasticity -- from that of synapse selection -- addressing. Explicit studies of memory addressing may find applications not only in artificial intelligence but also in biology. In artificial networks, for example, selectively updating memories in large language models can help improve user privacy and security. In biological ones, understanding memory addressing can help with health outcomes and treating memory-based illnesses such as Alzheimers or PTSD.
Subjects
Files
- Tyulmankov_columbia_0054D_18235.pdf application/pdf 4.36 MB Download File
More About This Work
- Academic Units
- Neurobiology and Behavior
- Thesis Advisors
- Abbott, Larry
- Degree
- Ph.D., Columbia University
- Published Here
- December 27, 2023