What happens when we delete a device?
By the rediscovery rate of a deleted devive, it suggests that not everything is actually deleted or disassociated with the device. What is retained on our account for our monitor and that device?
1 Like
Having played with the API calls quite a bit, I think I can answer this:
Deleted devices are basically flagged as inactive on your account and hidden from the Sense app. As far as I know all details are retained in the background. This is usually a good practice in general in case a delete ever needs to be undone.
Also, merging two devices sets both of those as inactive and creates a third new device that references the other two. This also makes sense in case you ever decide to unmerge.
3 Likes
The restoration of a deleowoukd be great but isn’t possible for us and support tells me is t exactly that easy for them either. They didn’t go into detail about it.
I can see where the data would need to be retained at the server level in database and used for detections on other monitors but don’t agree it’s a food idea for the monitor. It makes that data have a permanent home and the data is not “freed up” to be used for an entirely different device. It seems it’s like a windows delete in a way as but is not written over at some point after the indexing is removed
I don’t think any of the associated info is retained on your monitor once it is deleted / flagged inactive. Just retained out In the cloud.
1 Like
Is this something you’ve received a solid answer about from Sense?
It’s what I’m wanting to know because it has appeared to me that the “inactive” part is true but because of the rediscoveries, how quickly they happen and the closeness in resemblance, that something was retained. I’m not sure why it’s been so difficult to get this answer directly from Sense.
Not based on any direct answer from Sense, but on @brbeaird’s experiences with the API and my experiences with export, plus what I know about machine learning. Once a model “fits” a Device, that Device is typically going to be “found” much more quickly the second and subsequent times. Kind of like a combination lock - the first time you don’t know any of the parameters and you have to “unlock” the signatures by guided trial and error (in neural networks, guided trial and error is called gradient descent), but eventually you figure out all the parameters to detect a device and that becomes a model. Even after the Device association is deleted, the Sense cloud retains the model / combination of parameters to detect that Device.
1 Like
I must just not understand enough yet. I’m looking at it as putting together a puzzle. You have an area to fill and placing the correct pieces in where they belong while also going to see if a piece fits, realizing it does t and placing it to the side. When all the pieces are found and matched to where they belong, the “model” or “detection” is made.
That’s looking at the model as a whole where all the pieces are found. My thinking is when there are still missing pieces but this area of puzzle has been associated with a device, has any information at all been retained for later use. Like putting the puzzle away or taking a break. It’s still there but not yet completed.
Now you ha e me learning about backpropagation
1 Like
Many, many cycles (actually epochs) of back-propagation via gradient descent develop the combination of parameters to unlock detections.
1 Like
So I’m assuming when Sense started it was largely unsupervised learning due to a small base and lack of data. As more monitors were distributed and continue to be, now supervised learning is started.
The data collected during unsupervised learning and the data from supervised seems to cause a rub.
Maybe the data is quite similar but not exactly the same. Is it possible that they are able to use both methods simultaneously? Is there an interference and that why so many early adopters have such poor detection numbers even today? This also drives the question about what data is held onto. I hope that made sense.
I speculate that Sense bootstrapped their learning and models with a few houses (employees’) plus their lab doing supervised / reinforcement learning with some kind of proprietary feedback source (MIT, where the Sense founders came from, did disaggregation research research using their own laboratory version of smartplugs). From then on they probably kept adding devices of interest to their lab, and select alpha homes, while using unsupervised learning for the broader customer base.
1 Like
Now there’s a funny thought. Makes me wonder what the devices of interest were for MIT nerds. I can see Sense detecting some laser setup but missing my air handler.
MIT “Nerds” is a compliment, nothing derogatory
MIT’s work was done with real houses and laboratory-style metering equipment. Interesting read and fun to see how far we have come with the Sense monitor and smartplugs.
1 Like
I’ll be publishing the April vid later this week. In the meantime, keep the questions for next month coming!
1 Like
Here’s April’s video:
@samwooly1, I fear I’m spoiling you by having two of your questions answered in here… And naturally, the questions you asked for May are also great candidates!
7 Likes
I don’t take it as spoiling at all @RyanAtSense.
It a huge compliment to have our questions answered and the ones I ask are sure not just mine. They are the product of the dialogue and interaction with all the community members that I’m constantly learning so much from.
When I read all the questions in this thread, I wonder how difficult it would be to choose a few, they are all so good.
I really like the feeling of meeting and interacting with Sense employees.
Had a problem earlier could t finish:
Thanks @RyanAtSense and Sense for the really great answers this month. It’s really helped with a solid understanding of many things, far beyond the original questions.
1 Like
Another great video, I’m really enjoying these.
2 Likes
Thanks! Production value should continue to go up as well. Offloading some outdated gear and getting some nicer stuff 
. My pre-Sense life included teaching video production at a university, so this sort of thing is near and dear to my heart. I also love getting more folks than just me to answer all your questions. 
5 Likes
Always on continues to be a problem for a lot of us. The numbers we are shown are much higher than they should be or they grow without explanation or changes happening in our homes.
How about EXACT details about how Always On is calculated?
I’ve read the blog page that went up recently. It tells more than before but is still lacking.
2 Likes
Agree, Always On is a problem since the last update. My AO gone up considerably and it appears it’s consumed my pool pump which isn’t always on. I rarely see Other anymore.
Please keep the Always On discussion to the relevant threads.
1 Like