Multi-modal optimization techniques for improving qualitative features of biophysical neural models

Abstract

Event Abstract Back to Event Multi-modal optimization techniques for improving qualitative features of biophysical neural models Robert Clewley1, 2 and Mirza Dobric2* 1 Georgia State University, Neuroscience Institute, United States 2 Georgia State University, Department of Mathematics and Statistics, United States We present a novel computational technique that enables more efficient optimization of qualitative features in biophysical neural models. In particular, we extend the idea of multiple objective optimization for a single comparison modality (such as comparison of individual voltage traces) [1] into multiple modalities. For instance, features may be defined in terms of measurements from different experimental scenarios, and might include not only action potential spike shape characteristics (Cf. [1, 2]), but also complex features such as the frequency response curve (as a function of injected current), and phase response curve (as a function of perturbation time). The benefit of multiple feature modalities is to provide additional ways to distinguish solutions that might otherwise appear similarly fit in the restrictive view of a single modality. This is often due to relative insensitivities of some parameters to features in a single modality, or parameter co-variation/modulation [2], which creates an ill-posed optimization problem (poor gradients, local minima, or multiple global minima in the fitness landscape). Thus, for any given parameter choice, an evaluation of model fitness might involve multiple test scenarios to be executed. We use an implementation in the PyDSTool software [3] to demonstrate the ease with which this multi-modal approach can be prepared and embedded in conventional gradient-based or global optimization algorithms. We show that the technique improves detailed models with many parameters over heterogeneous data sets that contain noise and variability in underlying conditions, for which methods based on single modalities fail.