6 ms·

'It is about the relevant variables for deciding Y not being present in X.' Could you explain a bit more about this? Can I interpret that as 'X contains some si

'It is about the relevant variables for deciding Y not being present in X.' Could you explain a bit more about this? Can I interpret that as 'X contains some signals that indicate a target Y'? If that's the case then it still isn't that different from classical statistical inference, am I wrong? There are many problems that we can hardly be sure there are relevant signals in X when we are studying the target Y (e.g Biology/Physics), correct?

mjburgess 6y ago
I think the key point here is that ML is *automated* inference. In the sciences "inference to a model" is a process, not a dataset, called experimentation. The scientific method is the sequencing of relevant datasets, each of the latter "ruling out or in" the former. There are many issues with ML systems, fundamentally, that the "automation" involved is not an automation of science (/"learning proper") -- but the automation of function aproximation. A fatal issue is then that the datasets we are asking to be automatically aproximated are radically incomplete, and there is no "scientific method" within the ML system to appeal to. Eg., we are trying to get self-driving cars with devices (cars) which cannot estimate the intentions (mental states) of pedestrians (ie., a missing variable within the PD but not in X). And very profoundly, have even no means to acquire this data, or to temper their "decisions" via being aware of its absense.
- dumb1224 6y ago
  I take your point. However I do think the 'automation' process you described sounds like automated feature selection or the general idea of regression. 'A fatal issue is then that the datasets we are asking to be automatically aproximated are radically incomplete,' I couldn't agree more with that. For example, in my field of research a hypothesis is formed and can only be evaluated through a set of genomic profilings and clinical assessments. However no ML methods in the world could tell us whether these evidence are adequate to figure out the causal factors or there is any relevance of the dataset at all. Comparing with your self-driving car example, the error in the system and the cause are at least human-interpretable. That being said, it is a common theme in research and publications most people are aware of these caveats and limitations.

mjburgess 6y ago

I think the key point here is that ML is *automated* inference. In the sciences "inference to a model" is a process, not a dataset, called experimentation. The scientific method is the sequencing of relevant datasets, each of the latter "ruling out or in" the former. There are many issues with ML systems, fundamentally, that the "automation" involved is not an automation of science (/"learning proper") -- but the automation of function aproximation. A fatal issue is then that the datasets we are asking to be automatically aproximated are radically incomplete, and there is no "scientific method" within the ML system to appeal to. Eg., we are trying to get self-driving cars with devices (cars) which cannot estimate the intentions (mental states) of pedestrians (ie., a missing variable within the PD but not in X). And very profoundly, have even no means to acquire this data, or to temper their "decisions" via being aware of its absense.

dumb1224 6y ago
I take your point. However I do think the 'automation' process you described sounds like automated feature selection or the general idea of regression. 'A fatal issue is then that the datasets we are asking to be automatically aproximated are radically incomplete,' I couldn't agree more with that. For example, in my field of research a hypothesis is formed and can only be evaluated through a set of genomic profilings and clinical assessments. However no ML methods in the world could tell us whether these evidence are adequate to figure out the causal factors or there is any relevance of the dataset at all. Comparing with your self-driving car example, the error in the system and the cause are at least human-interpretable. That being said, it is a common theme in research and publications most people are aware of these caveats and limitations.