Thread: Chat gtp rem diction ideas using Lora scripts.

Hey I work with building machine learning models and this is a great idea and it's definitely possible to apply some version of this.

The way I see it there are two approaches how this could work. With one you would use an large language model like you're suggesting. In that instance I would say to either finetune one of the many existing open source models on the EEG REM data or train a LoRA if not finetune. But like on the numbers, I would not go for multi-modality here - stick with tokens.

And then the other option which could also work really well would be to create a image classifier on the graphing data! You would create a number of classes of states in the form of images. So one class is 'this graphing image means they're almost certainly in REM!" and one class can say 'this person is currently experiencing deep sleep' and so on. And then train the model on those classees. Then when it comes to inferencing in real time you would have the model continually read the current state of the graph for whoever is sleeping and it will constantly spit out it's confidence every moment of the state of each class.

For example you could get as many read outs a second of something like:

Patient's current state (confidence):

Deep sleep (87%)
REM sleep (02%)
etc for how many classes you want to track.

I have the hardware and experience to build such systems, especially image classifiers. I don't know much about EEG machines and whatnot though!

Just want to add that I looked into this more. I found a device called the Muse S which is a headband that can be worn at night and it tracks EEG and heart rate data and more. It does not have any dreaming functionality, but it does had an SDK (software development kit)! Which would give a developer access to read all of the live data and develop custom applications. It has Bluetooth connection which is ideal as no cables would have to be hooked up while sleeping.

I submitted a request for access to the SDK. I'm going to look into if the data coming out of this is reliable and how good the REM detection really is. If that looks good and my SDK request is approved I'm going to get one and build a model to try out your idea pointofbeing.