next up previous
Next: Experimental Conditions Up: Feature extraction through LOCOCODE Previous: Low-Complexity Autoassociators


Outline. Section 4.1 provides an overview of the experimental conditions. Section 4.2 uses simple artificial tasks to show how various lococode types (factorial, local, sparse, feature detector-based) depend on input/output properties. The visual coding experiments are divided into two sections: Section 4.3 deals with artificial bars, Section 4.4 with real world images. In Section 4.3 the ``true'' causes of the input data are known, and we show that LOCOCODE learns to represent them optimally (PCA and ICA do not). In Section 4.4 it generates plausible feature detectors. Finally, in Section 4.5 LOCOCODE is used as a preprocessor for speech data fed into standard backpropagation classifier. This provokes significant performance improvement.


Juergen Schmidhuber 2003-02-13

Back to Independent Component Analysis page.