BECCA 0.5.0 release

[Brandon Rohrer]

Hi all,

I finally completed work on a new version of BECCA. All the major algorithms have been reworked, including ziptie (a signal clustering method), daisychain (a temporal sequence construction method), and a novel reinforcement learning approach. Together, they perform better than any earlier incarnation on the benchmark tasks. They also perform a new visual search task, find_square, which finds a target against the background of a larger image. I updated the documentation as well. A lot of it I moved to the github project wiki. 

https://github.com/matt2000/becca/commit/122e326fa0bf270efcfbaaf314e5567143446c53
https://github.com/brohrer/becca/releases/tag/0.5.0

https://github.com/brohrer/becca_world_find_square/releases/tag/0.5.0

https://github.com/brohrer/becca/blob/master/doc/latex_files/how_it_works.pdf

Cheers!

Brandon

[SeH]

Thanks. The document helps make sense of the source code.  Would it be possible to indicate in the equations which variables refer to vectors, or when an equation is applied to each element of a vector?  I'm curious how large, in # of elements, each of the data flow links in the system are.  Maybe there's a way to estimate their eventual size at each stage, at least in proportion to the input/output space size.  I imagine a visualization of the entire system while it's running to make sure it's working right, and for experimenting with tweaking control parameters.

--
You received this message because you are subscribed to the Google Groups "BECCA_users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to becca_users...@googlegroups.com.
For more options, visit https://groups.google.com/groups/opt_out.

[Brandon Rohrer]

SeH,

Thanks. The equations in the paper don't describe exactly the calculations in the code--they just represent the important computations to illustrate the operation of the algorithms. It might be worthwhile to do a more formal documentation of the math at some point, but right now, it is still changing quite a bit from version to version, so I didn't invest the time. If I were going to invest it though, the first thing I would do is go to the code and put in a bunch of print xxx.shape lines.

Your idea of a visualization of the internal state of the system is a fantastic one. I did a little of that in the 'watch' and 'listen' videos (http://www.youtube.com/watch?v=GgWb0_ZoHXo&list=PLF861CC4C40439EEB&feature=share, and http://www.youtube.com/watch?v=EV7Rg1Jono4&list=PLF861CC4C40439EEB&feature=share&index=4), but it wasn't complete. Again, that's probably something I'll invest in once the algorithms stop changing so much from version to version. But at this point I couldn't guess when that will be.

Brandon

For more options, visit https://groups.google.com/d/optout.