Community Questions / Topics
I would like to hear Jeff's, Subutai's ideas on whole pattern stability given partial pattern input (along the lines of what has been a hot topic as discussed by Jake, Fergal, Rob, Chetan, Subutai, Feixdong, and myself) ? More specifically, how is 1234ABCD known/stabilized on in a previously trained model, given input beginning with "A" of ...ABCD?
- Given prior training for ABCDEFGH - when starting a fresh input at say D, how do the upper regions stabilize on the ABCDEFGH when all they will see is ...*DEFGH?
Given two patterns, ABC and ABCDEF - starting on ABC, how do you get it to discern whether it needs to keep matching and extend the recognition to ABCDEF when it may not be desired?
The second was brought up by Jake, the first Rob,Fergal and I have been arguing about since the dawn of humanity
Sensory Motor Work
I would like to hear about the new sensory-motor work and what to expect in the nupic.research repository.
- Matt Keith
If possible as well, I have a series of questions related to swarming in Nupic, or more generally on utilizing adaptive/evolutionary optimization techniques (particle swarm optimization being one example of these techniques) to determine the best model parameters for a region of an HTM network.
What was the reasoning for choosing PSO as the technique to determine the model parameters as opposed to other adaptive/evolutionary optimization techniques such as genetic algorithms, simulated annealing, tabu search, genetic programming, etc...?
Just out of curiousity has anyone had the opportunity to experiment with some of these other techniques (that I have mentioned above) in determining the best model parameters for a region?
Are there generic model parameters (numeric values) that either through experimentation, common sense, or neuroscience knowledge that you have found that work well regardless of the raw input data, or the application domain in which HTM is being applied to?
Initial response from Scott:
I am not sure what the original decision for PSO was but it is one of a set of algorithms that fit our specific problem. Within that set, PSO seems to have the best results:
- Some practical questions I ran into during te hackathon.
- How to handle false positives in processing big data sets.
- How to best handle separation of runs(tracks).