Testing predictive properties of efficient coding models with synthetic signals modulated in frequency

Testing the accuracy of theoretical models requires a priori knowledge of the structural and functional levels of biological systems organization. This task involves a computational complexity, where a certain level of abstraction is required. Herein we propose a simple framework to test predictive properties of probabilistic models adapted to maximize statistical independence. The proposed framework is motivated by the idea that biological systems are largely biased to the statistics of the signal to which they are exposed. To take these statistical properties into account, we use synthetic signals modulated by a bank of linear filters. To show that is possible to measure the variations between expected (ground truth) and estimate responses, we use a standard independent component algorithm as sparse code network. Our simple, but tractable framework suggests that theoretical models are likely to have predictive dispersions with interquartile (range) error of 4.78% and range varying from 3.26% to 23.89%.