Information-theoretic metric learning: 2-D linear projections of neural data for visualization

Area of Science:

• Neuroscience
• Computational Neuroscience
• Data Science

Background:

• Intracortical neural recordings generate high-dimensional data, hindering visual inspection of condition-specific responses.
• Dimensionality reduction is crucial for evaluating information content in neural data related to behavior or stimuli.

Purpose of the Study:

• To develop a method for projecting high-dimensional neural responses into a low-dimensional space.
• To create a metric that preserves class-relevant information for distinguishing between conditions.
• To visually evaluate the information carried by neural responses about different experimental conditions.

Main Methods:

• Linear projection to a 2-D space that parametrizes a metric between neural responses.
• Information-theoretic optimization to maximize information between projected data and class labels.
• Application to neural firing rates (macaque motor cortex) and local field potentials (rat somatosensory cortex).

Main Results:

• The projection method successfully reduces data dimensionality while preserving class-relevant information.
• Projected data points maintained the natural topology of experimental conditions (e.g., reaching targets, tactile sites).
• Utilizing the learned metric improved nearest-neighbor classification rates compared to original data.

Conclusions:

• The developed projection and metric effectively visualize and analyze neural responses.
• The method enhances the ability to distinguish between different experimental conditions using neural data.
• This approach offers a powerful tool for understanding neural coding in complex datasets.