Dimensionality reduction beyond neural subspaces with slice tensor component analysis.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Data Analysis

Background:

• Large-scale neural recordings are often analyzed using low-dimensional models focusing on neuronal coactivation.
• Existing models may overlook complex, higher-dimensional neural structures like sequences or evolving latent spaces.
• Task-relevant neural variability might exist in distinct, co-occurring 'covariability classes' over time or trials.

Purpose of the Study:

• To introduce a novel unsupervised dimensionality reduction technique for neural data tensors.
• To develop a method capable of demixing distinct covariability classes within neural activity.
• To extend the understanding of neural population activity beyond fixed low-dimensional subspaces.

Main Methods:

• Development of slice tensor component analysis (sliceTCA), an unsupervised method for neural data tensors.
• Application of sliceTCA to analyze neural activity patterns.
• Comparison of sliceTCA performance against traditional dimensionality reduction methods.

Main Results:

• Slice tensor component analysis (sliceTCA) effectively demixes distinct covariability classes in neural data.
• sliceTCA captures more task-relevant neural structure using fewer components compared to traditional methods.
• Demonstrated efficacy across diverse datasets, including primate motor cortex and mouse multiregion recordings.

Conclusions:

• Neural variability can be organized into multiple, higher-dimensional covariability classes.
• sliceTCA provides a powerful tool for uncovering complex latent structures in neural population activity.
• This framework expands the classic view of low-dimensional neural dynamics to include richer, higher-dimensional representations.