How often do the models update?

[kevin1]

@ramon,

The two things I have heard from Sense about model deployment over the years are:

1. A new model for a device in your household is deployed when Sense has enough “confidence” in the model reliably and unambiguously detecting a device it has identified. But how that is measured and what the threshold is is confidential, though it’s gotta be some probabilistic confidence measure.

2. The models are deployed via an automated process. There’s no human involved in looking at your inventory, and or initiating the push, though I’m sure some human interaction is involved in planning and maintaining “training cycles” for different models. And I would suppose that there are cases with new evolving models, where as Sense data scientist might deploy a model to some customer monitors just to check its behavior in-system.

I would guess that things slow down for the same reason I speculated here:

Minimum Consumption Threshold for Detection?

I’m not sure Sense is trying to find the “big fish” first. My view is that the issue with many of the small transitions is their lack of ubiquity - too many are similar and undistinguished in their “features” that Sense uses to identify unique fingerprints. Just to put a “face” on it, here’s a still from a Sense video back in 2018 that explained a little about their identification / categorization. It shows one slice of multi dimensional clustering of “features” used as input to identify dev…

Sense moves through the population of “interesting transitions”, building models and categorizing the easiest to unambiguously determine first. Those are the ones that happened with the greatest frequency of occurrence and greatest differentiation. In my mind, that means well-defined clustering in the 17+ dimension “feature space”. The greatest enemies of “detection” are probably:

• Device on/off transitions that don’t pass the Sense monitor “transition filter” - for instance EV charging, mini-splits, and electronics power supplies. Sense has other ways for dealing with some of these.
• Device transitions that are passed by the monitor, but form a big blobby cluster in the feature space that is too big to be one device and too undifferentiated to separate.
• Two or more devices with nearly the same feature space. I have some heating elements in my house that get confused frequently. Several in-floor heaters of about the same wattage plus my dryer heating element on normal dry.