Call To Arms
Earlier in the development of HTM.java, we had a call to arms regarding the development of encoders and this went very very well! We had a lot of people step up and take on the development of encoders and I was amazed at the turn out and the outcome.
I'd like to do this again.
Currently there is no way to save and reload a given network. This means we need to be able to serialize a Network and then deserialize it (reload it). Currently, the Python version of NuPIC uses Cap 'n Proto (sp?) to serialize their models and components - we should take a look at that and see if its appropriate for HTM.java? Saving and Reloading models or Networks may require an emphasis on speed, as the current thrust in NuPIC is to be able to swap models in memory very quickly. We should investigate the utility in this and align with the direction NuPIC is heading with this.
Once we decide on a serialization scheme, we need to divvy up the algorithms - each developer can take one or more algorithms and be responsible for its serialization. I have had an emphasis on having no state in the algorithms so the TemporalMemory and the SpatialPooler will need no serialization :-). However the Connections class will and so will the Network components - we have to see. Once the scheme is decided there will be lots of opportunities for developers to contribute.
So first I call on an enterprising developer or developers to research serialization schemes and generate a discussion to determine the path we should take...
If you would like to discuss this and inquire into whether this is something you might want to contribute, please drop in here:
...and have a chat with us!
I look forward to working with you all!
David (cogmission)