Fast and robust fixed-point algorithms for independent component analysis
IEEE Transactions on Neural Networks
| Hi-index | 0.00 |
It is known that independent component analysis (ICA) generates filters that are similar to the receptive fields of primary visual cortex (V1) cells. However, ICA fails to yield the frequency tuning exhibited by V1 receptive fields. This work analysis how the shape of IC filters depend on second-order statistics of the input data. Specifically, we show theoretically and through experimentation how the structure of IC filters change with second-order statistics and different types of data preprocessing. Here, we preprocess natural scenes according to four conditions: whitening, pseudo-whitening, local-whitening and high-passfiltering. As results, we show that the filter structure is strongly modulated by the inverse of the covariance of the input signal. However, the distribution of size in frequency domain are similarly biased for all preprocessing conditions.