Boston University
Contributions to marked point process models for neural spiking activity
Abstract
dc:description.abstractIn the past few decades, there has been a massive expansion in our ability to record neural activity: we can now record from many more neurons, across multiple brain areas, and at multiple spatial and temporal scales. At the same time, the technical advances in neuroscience experiments keep raising new challenges for how to analyze the resulting neural data. Large-scale and complicated data require a more powerful and flexible model structure, while current data analysis problems require increased accuracy and reduced computational complexity. Point process filters have been successfully applied to decode neural signals and track neural dynamics. More recently, a clusterless decoding approach based on non-parametric marked point processes has been developed, leveraging spike waveform features to characterize the relationship between a covariate of interest (e.g., a rat’s location on a track) and neural activity. This method, which incorporates waveform characteristics from all spikes and bypasses the need for spike sorting, has been shown to outperform traditional spike-sorting techniques. However, non-parametric joint mark intensity models are typically built using sums of kernel functions, whose computational complexity grows with the number of spikes and the dimensionality of the mark process. In this work, we explored more efficient clusterless joint mark intensity model structures. First, we proposed subsampling the spikes used to construct non-parametric joint mark intensity models. We proposed several subsampling strategies to select the subset of spikes that are used to build the intensity model. We applied the subsampled models to a real neural spiking dataset and showed the resulting estimates to be nearly as accurate as models based on the full dataset while reducing the computational cost substantially. Our results provide insight into the distinct contributions of low- and high-amplitude spikes in the neural model. We find that both types of spikes carry essential yet largely distinct information, while the total information across all spikes is highly redundant. In addition, our subsampling methods provide an effective approach for identifying a representative subset of spikes that retains nearly all the information necessary for accurate decoding. Second, we built a parametric mixture of Gaussian joint mark intensity model (MoGJMI) and developed an EM algorithm to estimate the Gaussian mixture parameters by introducing the latent variables that represent the identification and classification of spikes in different mixture clusters. The MoGJMI uses a small number of clusters to represent the place field structures and spike waveform distribution of the neural population with high accuracy and efficiency. The simple model structure allows MoGJMI to incorporate the history dependence of the clusters and interaction effects among different clusters. This has the potential to capture the complex history dependence structure and relationships known to exist in networks of neurons.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy
- Level thesis:degree_level
- doctoral
- Discipline thesis:degree_discipline
- Mathematics & Statistics
- Grantor
- Boston University
- Year dc:date.issued
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Li, Jingjing
- Advisor dc:contributor.advisor
-
- Eden, Uri
Subjects
dc:subject × 1Rights
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/2144/53374