Identifying functional bases for multidimensional neural computations.

Area of Science:

• Computational neuroscience
• Neural coding
• Systems neuroscience

Background:

• Current dimensionality reduction methods identify relevant neural subspaces but struggle to interpret the basis within them.
• Interpreting nonlinear neural computations across multiple dimensions simultaneously is challenging without simplifying assumptions.
• Relating predictive models of neural activity to mechanistic descriptions of neural circuitry remains a key challenge.

Purpose of the Study:

• To develop a method for transforming neural data into a "functional basis" that simplifies the description of neural computations.
• To demonstrate how this functional basis facilitates the interpretation of neural computations and the development of mechanistic models.
• To illustrate the utility of functional basis transformation using logical functions and real neural data.

Main Methods:

• Dimensionality reduction techniques (e.g., spike-triggered covariance) were used to identify relevant subspaces.
• A maximum likelihood approach was employed to find the functional basis within the relevant subspace.
• Simulated neurons and retinal ganglion cell recordings were used for illustration.

Main Results:

• Transforming to a functional basis simplifies the description of nonlinear neural computations, often with fewer parameters.
• Standard dimensionality reduction methods may yield difficult-to-interpret dimensions that do not align with a functional basis.
• The identified functional features are unique, nonorthogonal, and improve the link between computational and mechanistic models.

Conclusions:

• The functional basis provides a simplified and interpretable representation of neural computations.
• This approach enhances the ability to connect abstract computational models with underlying neural circuit mechanisms.
• The method offers a powerful tool for elucidating neural computation in various biological systems.