Sorry Matt, that went out before I got your mail. Moving this over to another list.
>From Hackers: ---------------- Quick Question: I'm working on the Network API for htm.java and I was wondering what the normal usage pattern is for input data consisting of multiple fields. Obviously, if the input data consists of different types, different encoders must be used. Does this also mean that each "type" corresponds to a unique SP/TM pair ? And what of Anomaly instances, do we need a different one of those for every distinct "type" of data input? Does the same thing go for Classifiers? Maybe that was a little more than a "quick question" :) Thanks, David ------------- >From Scott: I'm not sure if this is what you mean but the MultiEncoder is responsible for taking multiple fields (same or different types) and creating a single SDR. That is passed to the SP and from there on you just have a single SDR throughout. The anomaly score is computed on the entire TM. The CLA classifier, however, predicts a specific input field so you have to pass that field from the encoders to the CLA classifier. But the CLA classifier isn't very cleanly implemented and you currently have to manually call a custom compute method on it since it isn't set up to work entirely within the network api. I think we can probably fix that though. -------------- Scott, Thanks for the prompt response. How about Anomaly instance usage? I would assume a single instance of an Anomaly tracks the anomaly scoring for a single unique field, correct? So then, there must be one per field entry type no? David On Thu, Feb 12, 2015 at 12:53 PM, cogmission <[email protected]> wrote: > Scott, > > Thanks for the prompt response. > How about Anomaly instance usage? I would assume a single instance of an > Anomaly tracks the anomaly scoring for a single unique field, correct? So > then, there must be one per field entry type no? > > David > > On Thu, Feb 12, 2015 at 12:50 PM, Scott Purdy <[email protected]> wrote: > >> I'm not sure if this is what you mean but the MultiEncoder is responsible >> for taking multiple fields (same or different types) and creating a single >> SDR. That is passed to the SP and from there on you just have a single SDR >> throughout. The anomaly score is computed on the entire TM. >> >> The CLA classifier, however, predicts a specific input field so you have >> to pass that field from the encoders to the CLA classifier. But the CLA >> classifier isn't very cleanly implemented and you currently have to >> manually call a custom compute method on it since it isn't set up to work >> entirely within the network api. I think we can probably fix that though. >> >> On Thu, Feb 12, 2015 at 10:39 AM, cogmission <[email protected]> >> wrote: >> >>> Quick Question: >>> >>> I'm working on the Network API for htm.java and I was wondering what the >>> normal usage pattern is for input data consisting of multiple fields. >>> Obviously, if the input data consists of different types, different >>> encoders must be used. Does this also mean that each "type" corresponds to >>> a unique SP/TM pair ? And what of Anomaly instances, do we need a different >>> one of those for every distinct "type" of data input? Does the same thing >>> go for Classifiers? >>> >>> Maybe that was a little more than a "quick question" :) >>> >>> Thanks, >>> David >>> >>> -- >>> *We find it hard to hear what another is saying because of how loudly >>> "who one is", speaks...* >>> >> >> > > > -- > *We find it hard to hear what another is saying because of how loudly "who > one is", speaks...* > -- *We find it hard to hear what another is saying because of how loudly "who one is", speaks...*
