Hi Marek, I totally agree with you about the need for hierarchies in NLP, an inherently hierarchical problem. When working on Linguist, I found that it can learn sequences of letters decently, especially when you give it repeated training, but it doesn't seem to be able to generalize well (or even at all) to new data.
Thus, I've also been interested in adding in a hierarchy, but haven't yet gotten a chance to properly think about it. I'm glad someone is looking into it. Please do keep us posted on your efforts! Thanks, Chetan On Wed, Oct 16, 2013 at 4:05 AM, Marek Otahal <[email protected]> wrote: > Hi all, > > I'm implementing a hierarchical CLA by connecting the regions (implemented > in ROS). > > The easiest way seems to be to run multiple OPF clients and chain output > of one to input of some other. Recently we got PassThru encoder merged in > (thanks Keithcom!) so this part is ok. > > My question is, can I make an OPF model that uses some encoder (to > transform real world data) but outputs an SDR? > > Eg., I want to encode letters as SDRCathegory (see linguist project), but > output a SDR so I can pipe it to other models with PassThru enc, and at the > end of the pipeline again feed in a SDR but this time decode back to > character. > > > Btw NLP sidestep: I was playing with the linguist project for text > generation, and I wonder why the performance is quite bad (compared to the > state-of-the art - deep NN) and I attribute it to the lack of abstraction > at higher levels of hierarchy - hence the ROS hierarchy! Do you have an > input how to increase performance there? > > Thanks for your help, > Cheers, > ~breznak > > -- > Marek Otahal :o) > > _______________________________________________ > nupic mailing list > [email protected] > http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org > >
_______________________________________________ nupic mailing list [email protected] http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
