The medical software that learned the worse the underlying condition is the better off you are with respect to pneumonia case has been rattling around my head for a few hours.
Do we have any concept of an intervention in machine learning? I am sort of gesturing at the Judea Pearl sense of the word, but in the end I really mean physical attempts to change things. So the basic question is, how does the machine learn when we have tried to change the outcome already?
Does an intervention have size or number? Like the difference between trying one thing, like taking an aspirin, or trying many things, like an aspirin, and bed rest, and antibiotics?
Do interventions have dimension? Like is there a negative intervention where we try to stop or reverse a process, and positive intervention where we try to sustain or enhance a process? Would we consider pneumonia interventions trying to sustain/​enhance lung function, or stop/​reverse disease progression? Presumably both in a suite of advanced care.
Does uniformity across data in terms of interventions predict the success or failure of machine learning approaches? Example: fails with different intervention levels in medicine; succeeds with unintervened sensor data like X-rays; also succeeds with aligning lasers in fusion reactions, which are in a deep well of interventions, but uniformly so.
The medical software that learned the worse the underlying condition is the better off you are with respect to pneumonia case has been rattling around my head for a few hours.
Do we have any concept of an intervention in machine learning? I am sort of gesturing at the Judea Pearl sense of the word, but in the end I really mean physical attempts to change things. So the basic question is, how does the machine learn when we have tried to change the outcome already?
Does an intervention have size or number? Like the difference between trying one thing, like taking an aspirin, or trying many things, like an aspirin, and bed rest, and antibiotics?
Do interventions have dimension? Like is there a negative intervention where we try to stop or reverse a process, and positive intervention where we try to sustain or enhance a process? Would we consider pneumonia interventions trying to sustain/​enhance lung function, or stop/​reverse disease progression? Presumably both in a suite of advanced care.
Does uniformity across data in terms of interventions predict the success or failure of machine learning approaches? Example: fails with different intervention levels in medicine; succeeds with unintervened sensor data like X-rays; also succeeds with aligning lasers in fusion reactions, which are in a deep well of interventions, but uniformly so.