Neuromodulation Stimulus Design With Data-Efficient Pseudoinverse Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for estimating pseudoinverses in neuromodulation require large amounts of data and are inefficient, particularly when dealing with non-invertible forward mappings and limited data availability, and they often fail to explore a large number of parameters effectively.
Innovation Solution
A novel pseudoinverse estimation system that jointly learns a restricted domain and the inverse mapping using regression techniques, minimizing data requirements by optimizing a weighted L2 loss to approximate an indicator function over the restricted domain.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If data-driven methods are used to estimate pseudoinverse, then the number of explorable parameters increases, but the data requirements increase
Solution Approach 1:
The patent inverts the traditional approach by directly estimating the pseudoinverse mapping from neural response to stimulus parameters, rather than estimating the forward mapping and inverting it numerically. This inversion approach allows direct regression from response to parameters, reducing data requirements while maintaining the ability to explore multiple parameters
Solution Approach 2:
The patent changes the parameter estimation approach by using a restricted domain with learned weights that approximate an indicator function. This allows the model to focus on a subset of parameter space that is most relevant for achieving desired neural responses, reducing the effective data requirements while maintaining versatility
2Measurement precision
If numerical inversion of forward model is used, then pseudoinverse can be estimated, but small errors in forward model cause inversion to blow up
Solution Approach 1:
The patent avoids numerical inversion by directly estimating the pseudoinverse through regression. This eliminates the instability associated with inverting approximate forward models, as the regression approach directly learns the mapping from responses to parameters without amplifying small errors
Solution Approach 2:
The patent uses a restricted domain approach with learned weights that act as a form of regularization. This beforehand constraint prevents the model from overfitting to noise or small errors in the data, cushioning against instability while maintaining estimation accuracy
3Measurement precision
If conditional density estimation methods are used, then pseudoinverse can be estimated, but more data is required compared to regression
Solution Approach 1:
The patent uses a direct inversion approach through regression, estimating the pseudoinverse by regressing stimulus parameters on neural responses. This is more data-efficient than conditional density estimation methods, which require sufficient data to estimate probability densities, while still achieving accurate pseudoinverse estimation
Solution Approach 2:
The patent changes the estimation approach from density-based to regression-based parameter estimation. By directly regressing parameters on responses with a restricted domain and learned weights, the method achieves comparable or superior performance with fewer data requirements
Data Source
AI summary
Disclosed herein is a novel pseudoinverse estimation system and method that adapts regression techniques to directly estimate one or more pseudoinverses of a neuromodulation pathway, thereby circumventing the need of inverting an estimated forward model. This is accomplished by the learning of a restricted domain that restricts the potential stimuli required to produce a desired neuro response.