Please read this thread. It discusses why it is not a feasible inclusion.
If we felt that manual training would a good use of your time, we would work on implementing it
Wow, that’s a tad bit condescending to say, especially when customers are declaring it to be worth their time.
Don’t take it out of context.
There’s posts here complaining about just the opposite on how they “have to scurry around” to I.D. discovered devices. There’s a thousand different personalities here which means a thousand different levels of expectations.
This beats everything on the market… If it discovers nothing, it’s still as good, but cheaper than TED, The Energy Detective.
I have both and couldn’t agree more, I actually have five different methods of energy monitoring and far prefer Sense.
@robertalangerhart, you are commenting on an ancient post. Ryan and Sense team have done an excellent job explaining why “human training” is not technically feasible, despite all the obvious enthusiasm and imagined viability.
First, if I’d wanted your opinion, I would have asked you. Second, cell phones can be trained to recognize a face, Google assistant can be trained to distinguish an individual by voice based on three samples, an electronic doorknob can be trained to recognize a finger print, yet somehow it is not “technically feasible” for Sense to be trained recognize a refrigerator or a garbage disposal. Whatever.
Robert maybe you should spend some more time reading the forum before initiating personal attacks on other forum members. Cherry picking a single developer’s quote doesn’t explain the work or thought process going into this product. This is a topic that has been beaten to death and Sense has addressed it in multiple posts so please put in some effort to read up on it.
@robertalangerhart, the kind of face “training” or the voice “training” you describe, is done in Sense as well, when you fill in the type, make and model of a new device it has identified. You can see the results when it offers up community-sourced options.
But there are 15 years of widespread university and private research that have gone into the precursor step in the training process - finding generic faces in photographs or identifying a string of words spoken in any voice. For reasons described earlier, untrained humans aren’t really even technically competent to tag the power waveforms accurately enough with ground truth information, to provide the annotated dataset for this kind of training.
Cut some slack. Google resources are huge. The sub-team that focuses on voice recognition for Google Assistant alone is more than of all bodies that make up Sense. Nuance built a foundation of computer speech recognition over 2 decades ago.
The industries you mentioned have standards required to accommodate those technologies. Fridge and garbage disposal companies don’t even know Sense is a thing.
Then explain how there are 1000 different OCR vendors out there recognizing text that wasn’t printed based on any “standard”. I’ve spent 20+ years in AI, and this is entirely do-able. I can train a missile to fly through unknown terrain to take out an enemy tank vs a friendly, yet you defend how it is too hard to identify a table lamp. #whatever
Common text that we all use every day and that OCR’s are trained to read come from different alphabets. Alphabets such as Latin, Greek, Arabic, Hebrew etc… the list is quite long.
These alphabets are a standard of which all scripts / fonts / representations of the alphabet are based on. When someone creates a new font/script or style it is based on the standard. If it wasn’t, we as humans, let alone machines, would not be able to read them.
Now lets take that standard alphabet and put a bunch of letters on top of each other.
Can the 1000 OCR vendors recognize the text in the below image?
I doubt it. Now as a human, with some context clues, you may be able to figure it out because you see more than the computer does, but just because you can see it and detect it doesn’t mean the computer can nor should be able to.
I don’t understand your hostility to a a team of people who are actually building the product. A team of people who have given countless well thought explanations of how their product works and their process for how it is supposed to work. Maybe there is a method to training an electrical detecting device, but its not this device. And I choose to believe that if training it were possible, and as simple as many people like to say it is, either a) Sense would be doing it already as why wouldn’t they want the best product they can, and or b) that there would be a competing product on the market that allows training and has a reliable track record.
If you don’t like the product, return it. If you have actual examples using your 20 years of AI knowledge of how they could program this and you have code and ML algorithms reference examples, then by all means share. If you don’t feel like sharing that knowledge, then you might as well keep your resume to yourself as it benefits no one other than just making you look like a troll.
Why don’t you just buy a Beagleboard with some A/D converters, hook up some CTs, rent some machine learning server capacity from AWS and develop your own LSTMs for identification. Seems quite trivial for an intellect like you…
Seriously dude, you seem not to know the history of ML-based image recognition by your allusions that it’s evolution was simple and fast. It took 6 years, plus a highly evolved and tagged dataset , plus fierce competition to get where we are in image recognition today.
Not defending Sense - we goad them enough for progress. But your comments are just arrogant, yet ignorant enough about ML, to trigger annoyance. I seriously want to see you tag some on and off-events with millisecond accuracy.
Guys, this thread is getting locked. Civil discussion about machine learning and what you wish we were doing differently is fine. This has dissolved into a mess of name-calling and is not what this forum is about. Feel free to continue these discussions in private messages. I encourage everyone to review our Community Guidelines: Sense Community Guidelines
Our feelings on this at the moment are here: Why can't you train Sense?
We’re constantly working on ways to get better user input and do not feel that a training mode is the best way to do that.
Wouldn’t it be nice to tun on a device (e.g. oven) with power meter view open & be able to tap the graph label associated with the meter spike and enter the device name/model…
I’ve just posted about the same:
Yes, a 1000x this. I understand Sense is hoping to parse readings and define appliances algorithmically, but adding a more robust crowd-sourced manual identification process is going to help reduce the loss of dissatisfied users (most due to lack of identifications or poor identifications - give them the opportunity to make up for current shortcomings with Sense via manual identifications) and add more data to the system to refine the algorithms. Users will of course make mistakes which will need to be addressed over time, but short-term I don’t imagine this would outweigh the benefits. Maybe sandbox user identifications created this way until Sense can validate them for sharing more broadly - satisfies all the customers complaining about lack of or incorrect identifications, and continues to feed data into the algorithm process.
Until they add some sort of feature which allows users to self identify appliances - the device is pretty much useless!
After having the unit for almost 6 months it has only identified a dozen or so appliances, only about 30% of my total energy load. If the unit allowed me to label items as they come on - I could shut everything in the house off, then turn appliances on one at a time and labeling them accordingly.
Until such a feature exists - I would not recommend spending the money for a “Sense”.
I agree. This feature should have been part of the MVP.
When I purchased the product I thought it already had it and notifications in the app itself lead you to believe that “Walk around, turning devices on and off, and check out their power usage.” is something you can save or log in the app.
That’s referring specifically to the Power Meter function of the app.
You can learn more about why such a training mode is not feasible here: Why can't you train Sense?