@Bitking: I tested LateralPooler for my VGA-Video, but it run too slowly even for C++ version. Moreover, it is very hard to understand or to say if the output is good. By DL, we can see and easily understand all learnt features.
I had some preliminary results which suggested it worked reasonable well, at least as well as the encoder with fixed bins repeated for every pixel. But I don’t have anything rigorous and reproducible to show you.
Visual data took too much time for my HTM to process so I switched the datasets I’m working with. At some point I do hope to use my new image encoder though, someday.