@article{Understanding:7504,
      recid = {7504},
      author = {Zhu, Yuqing},
      title = {Understanding Neocortical Dynamics and Computation through  Spiking Neural Network Modeling},
      publisher = {The University of Chicago},
      school = {Ph.D.},
      address = {2023-08},
      pages = {162},
      abstract = {Through the use of biofidelic spiking neural network  models (SNNs), this work offers mechanistic insights into  the relationship between neocortical structure, dynamics,  and computation. To set the stage, the topics of cortical  computation, ANNs as applied to neuroscience, SNNs as  applied to studying structure-function relationships in  neocortex, and the challenges of model interpretability are  introduced. In the first chapter, in the tradition of using  SNNs to address structure-function questions, we ask why  stable dynamics are possible in neocortex. Biofidelic SNNs  were created through a grid search for architectures that  yielded realistic spiking activity that was low-rate,  asynchronous, and near-critical. The maintenance of this  activity is linked to patterns of higher order coordination  of synaptic activity, and this coordination takes the form  of transitions in time between specific three-unit motifs.  These motifs summarize the way spikes traverse the  underlying synaptic topology. The second chapter turns its  focus to computation, which occurs in neocortex on a  substrate of stable activity that we studied in the first.  Specifically, this chapter covers why computation becomes  possible through specific synaptic changes and resulting  dynamic changes that occur through learning. After training  SNN models to perform an ethologically relevant task,  models come to selectively adjust firing rates in response  to the stimulus input. Excitatory and inhibitory  connectivity between input and recurrent layers changed in  accordance with this rate modulation. In particular,  recurrent inhibitory units which were tuned to one input  over the other strengthened their connections to recurrent  units of the opposite tuning. We conclude by discussing the  potential of task-trained SNNs for hypothesis generation  and testing in future research on neocortical computation.  Additional neocortical features that may be important for  computation are surveyed, and questions of model  interpretation are revisited in the context of these  results.},
      url = {http://knowledge.uchicago.edu/record/7504},
      doi = {https://doi.org/10.6082/uchicago.7504},
}