2.1. Connective Field (CF) modelling¶
2.1.1. Benchmarks for different modelling procedures in cortical connective field (CF) mapping.¶
Here we compare four flavors of population cortico-cortical connective field (CF) modeling. First, the former implementation in mrVista (using Matlab’s lscov) which implemented a grid search approach [1,2]. Without further optimization, the parameter space revealed by this implementation matches predefined grid value predictions. Second, the implementation part of the connective field branch of the prfpy Python package [3]. Using Python scikit-learn optimization functions (implemented as part of the prfpy package), the parameter space prfpy provides also converges to predefined grid, although at a much faster rate (thanks to the CPU parallelization tool joblib). Third, a custom version of more recent CF modelling approach (by So-Hyeon et al. (2024) [4]) that implements a derivative-free parameter (CF size) refinement approach (see here). Fourth, our custom Python implementation of automatic differentiation-powered gradient descent using TensorFlow and CUDA (partly inspired by the recent Python package braincoder [5]) to achieve highly efficient gradient descent (see here and here for a joint optimization approach that optimizes both CF size and position). To coordinate this effort, we rely on widely used population receptive field mapping tools [6,7] and high-field 7T-MRI retinotopy data kindly made available by NeuroSpin.