Abstract
A neural implementation of the JADE algorithm, called nJADE, is developed which adaptively determines the mixing matrices to be jointly diagonalized with the JADE algorithm. This alleviates the problem of algebraically determining these mixing matrices which becomes a very tedious if not impossible undertaking with high dimensional data. The new learning rule uses higher-order neurons ...
Abstract
A neural implementation of the JADE algorithm, called nJADE, is developed which adaptively determines the mixing matrices to be jointly diagonalized with the JADE algorithm. This alleviates the problem of algebraically determining these mixing matrices which becomes a very tedious if not impossible undertaking with high dimensional data. The new learning rule uses higher-order neurons and generalizes Oja's PCA learning rule. As a test case the new nJADE algorithm is applied to high dimensional natural image ensembles to learn appropriate edge filter structures. Quantitative comparison concerning various filter characteristics is made with results obtained with a probabilistic ICA algorithm with kernel-based source density estimation.
Altmetric