An Experimental and Theoretical Study on Structural and Functional Plasticity in Cortical Neurons: Implications for Learning in Deep Neural Networks
Abstract
Project SummaryWe propose to elucidate inhibitory structural plasticity in relation to synaptic distribution,arbor location, and connection identity, as well as the location and dynamics of their excitatorycounterparts. Leveraging our experimental capability to distinguish synapses according to afferentsource, we will examine how inhibitory and excitatory synapses that differ in input source undergoplasticity. We will also study changes in synaptic strength (functional plasticity) based on the sizeof the post-synaptic scaffold. Our goal is to generate a theoretical framework for interpreting theexperimental data (inhibitory and excitatory synaptic placement delineated according to afferentsource, and their structural and functional remodeling), towards a coherent understanding ofsynaptic plasticity and resultant learning in single neurons. While biophysical models are currentlyused to describe the input/outputfunction of single neurons, we lack an efficient way to #teach#these biophysical neuron models to implement a computation (e.g. an XOR operation). We alsopoorly understand what makes a computational function easy or difficult for a neuron to learn.Here we will develop a novel learning rule that incorporates both changes in synaptic weights andin synaptic wiring, using deep neural network (DNN) models as neuronal surrogates. We will usethis knowledge to explore how neurons perform real-world tasks such as image classification.Specific Aims:Aim 1: To track the structural remodeling of inhibitory synapses, how it relates to theirafferent input specificity, and whether sensory manipulation differentially affects inhibitorysynapses dependent on their afferent source.Aim 2: To characterize the temporal dynamics of #top down# and #bottom-up# excitatoryafferent inputs (cortico-cortical and thalamo-cortical), and determine whether plasticityrules elicited by sensory deprivation differentially affects their distribution or dynamics.Aim 3: To extractfrom Aim 1 & 2 imaging datasets information on inhibitory and excitatorysynaptic size as a proxy for synaptic strength, adding an as yet unexplored functionaldimension to the structural computational cell modeling.Aim 4: To develop a detailed biophysical model for L2/3 cortical pyramidal cells based onthe synaptic mapping results from Aims 1-3.Aim 5. To develop a machine learning framework for understanding learning in a singleneuron that incorporates both functional and structural synaptic plasticity.In our prior ONR grant we developed a method to transform a biophysical model with a realisticnumber of synapses, experimentally determined, into an analogous DNN. Here we will use thismodel surrogate to faithfully represent L2/3 pyramidal model neurons developed in Aim 4, so thattheDNN will match Aim 1-3 experimental results. We will further incorporate novel deep learningarchitectures to increase the accuracy and efficiency of the model. We will then incorporate thedetails of synaptic connectivity and plasticity revealed by Aim 1-3 data and explore the scope andboundaries of our learning algorithm under these constraints. Lastly, we will incorporate thesingle-neuron-DNN into a multi-layer DNN and explore the computational capabilities of the latterwhen consisting of units that, on their own, are already #deep#. Overall, our novel learning ruleand the interplay between the biophysical and DNN model of the single neuron will enable, forthe first time, a systematic exploration of the computational capabilities of a single neuron acrossdiverse tasks and elucidate the synergistic role of functional and structural plasticity in achievingthese tasks, based on the experimental part of this study.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 24, 2024
- Source ID
- N000142412055
Entities
People
- Elly Nedivi
Organizations
- Massachusetts Institute of Technology
- Office of Naval Research
- United States Navy