In the ZCA whitening function, would it be a good idea to expose the epsilon as an argument (with default value)?
It seems that it's an important knob to control the effect of noise / high-frequency content on the whitening.
Here are screenshots demonstrating that. This is k-means on image patches. The first screenshot is with the default epsilon (1e-5), the second with a higher epsilon of 1e-3, the third with and epsilon of 1e-1. These are completely empirical observations of course. Here it seems that the second one yields the best basis.
In the ZCA whitening function, would it be a good idea to expose the epsilon as an argument (with default value)?
It seems that it's an important knob to control the effect of noise / high-frequency content on the whitening.
Here are screenshots demonstrating that. This is k-means on image patches. The first screenshot is with the default epsilon (1e-5), the second with a higher epsilon of 1e-3, the third with and epsilon of 1e-1. These are completely empirical observations of course. Here it seems that the second one yields the best basis.