Scaling down nupic model size

Engineering NuPIC
1

So I’m looking to add a ‘model-scale’ parameter to my function that generates nupic models. A set value of 1.0 would be a full-sized model, while a value of 0.5 would be a half-sized model, etc.

I want to make sure I adjust the SP and TM params in a valid way to achieve this, so here is the logic I have to do so:

SP

  • ‘column_count’: 2048 --> 1024

TM

  • ‘column_count’: 2048 --> 1024

  • ‘input_width’: 2048 --> 1024

  • ‘newSynapseCount’: 20 --> 10

  • ‘maxSynapsesPerSegment’: 32 --> 16

  • ‘maxSegmentsPerCell’: 128 --> 64

  • ‘minThreshold’: 9 --> 5

  • ‘activationThreshold’: 12 --> 6

Is there anything missing from this??

I’m using the ‘getScalarMetricWithTimeOfDayAnomalyParams’ to generate the params with RDSE encoders. Should the size of these encoders change as well? I assume that the smaller SP would map its fewer columns to the same-size encoder, resulting in bigger receptive fields for each column. Is that assumption correct?

Thanks as always :smile:

2 Likes
2

Could I get a quick thumbs up/down or re-direct @rhyolight?

The reason for the scaling is to try and reduce the memory use of each model, since I have n of them running at once. However I am skeptical that this scaling would reliably help – since I suspect that model memory-size depends more on the complexity & noise of the data than model scale. This is my hunch given the nature of HTM parameters values as sparse, simple (no heavy math) and accumulated from 0.

3

It seems ok to me Sam, but I have not experimented with this type of thing at all yet. So I have to also add :man_shrugging:.

2 Likes