Dimensionality reduction beyond neural subspaces with slice tensor component analysis
Background Introduction: Large-scale neural recording data can typically be described by patterns of co-activated neurons. However, the view of constraining neural activity variability to a fixed low-dimensional subspace may overlook higher-dimensional structures, such as fixed neural sequences or slowly evolving latent spaces. This study argues that task-related variability in neural data can also co-vary across trials or time, defining different “covariability classes” that may co-exist within the same dataset.
Research Motivation: Traditional dimensionality reduction methods (e.g., Principal Component Analysis (PCA)) typically capture only a single covariability class. To disentangle the mixed covariability classes, researchers developed a new unsupervised dimensionality reduction method called Slice Tensor Component Analysis (SliceTCA).
Research Method: SliceTCA is a novel tensor decomposition method based on slice rank. It decomposes the neural data tensor into three types of slice components: neuron slices, trial slices, and time slices, capturing different types of covariability. Unlike other dimensionality reduction methods, SliceTCA can simultaneously fit these three types of slice components, thereby disentangling the mixed covariability classes present in the data.
The researchers constructed toy models and recurrent neural network models to validate SliceTCA’s advantage in disentangling mixed covariability classes. They also proposed a standardized procedure for SliceTCA model selection, optimization, and visualization, and demonstrated SliceTCA’s application on three large-scale neural datasets.
Research Results: 1. Motor cortex activity data: SliceTCA found that trial slice components captured neural sequences related to movement dynamics, while time slice components captured information related to movement preparation.
Mouse motor task data: SliceTCA identified covariability related to task states and brain region specificity, exhibiting more interpretable neural representations.
Multi-area recording data: SliceTCA disentangled cross-area covariability types and identified components related to trial performance, population coding, and task stages.
Furthermore, the researchers provided a geometric interpretation of the different latent variable types corresponding to the slice types. Overall, SliceTCA captured more task-related covariability structures and represented neural data with fewer components, extending the classical low-dimensional neural population activity view.
Significance: SliceTCA offers a novel unsupervised dimensionality reduction method that can identify and disentangle multiple covariability classes mixed within the same dataset, revealing higher-dimensional latent structures. This method not only deepens the understanding of neural coding mechanisms but also facilitates the extraction of behavior-related information from large-scale neural data. By integrating multiple covariability classes into a unified framework, SliceTCA extends the classical views of latent variables and neural covariability. This study provides new theoretical foundations and practical tools for extracting behavior-related information from complex neural data.