"I would temper this with we are likely in the age of unmarked data and unstructured data. Less often are we asked to automate a known prediction and more often we are asked to cluster, characterize and segment wild data. In my opinion the hard problem in machine learning has moved from prediction to characterization."
I would say instead we are in the age of Discovery. The hard problem is discovering causes that give rise to coherency in large sets of real world data.
As to whether we are in the age of large computers, the classification is irrelevant. We are in the age of sufficiently large computers. There is sufficient power and scale to accomplish everything that the brain does with our current infrastructure. Is it highly inefficient on a watt/thought basis? Yes. But that's a secondary metric.
To address the concerns the author outlines:
Dealing with lack of training data - when you move to discovery labeled data is far less relevant, which should relieve research assistants who are currently being asked to label images as cat/not-cat.
Dealing with large sets of unstructured data - this is a processing problem for which the current infrastructure is sufficient.
IMO we have plenty of data, we don't need many more labels, and we have sufficient processing power. That leaves algorithmic work for the discovery of causes. (My bet is on systems similar to the kind of sparse distributed hierarchical models that Numenta is working on.)
One thing that I don't think most people have realized is how far past the quantity of data and processing power we need we already are.
If you were to take all the visual information the brain receives in the first 30 years of life (plenty of time to become intelligent) you'd arrive at a number around 1.4 petabytes. This is hardly a large number compared to the amount of information processed by YouTube on a monthly basis. Hearing, taste, and smell all add in a fraction more, and touch (well I don't have any good numbers there yet).
The hard part now is going to be convincing machine learning researchers to actually expose their algorithms to a swath of data similar to the inputs we humans get, over sufficient time, so as to allow for the development of a robust hiearchy of stored causes (aka understanding.)
I was hoping nobody would outthink me and ask that. I am going equivocate a bit. Now that we have kernel methods (for interactions) and graphical models (for dependencies) you could claim prediction is "solved" in that if you guess the right feature design (no small task) current optimization methods are up to the task of prediction. Of course some better method could come along and we would be forced to say the current level of prediction turned out to be inadequate. Clustering and characterization are indeed where a lot of the current applications lie and can range from fascinating to unpleasant (depending on details of the problem domain).
I would say instead we are in the age of Discovery. The hard problem is discovering causes that give rise to coherency in large sets of real world data.
As to whether we are in the age of large computers, the classification is irrelevant. We are in the age of sufficiently large computers. There is sufficient power and scale to accomplish everything that the brain does with our current infrastructure. Is it highly inefficient on a watt/thought basis? Yes. But that's a secondary metric.
To address the concerns the author outlines:
Dealing with lack of training data - when you move to discovery labeled data is far less relevant, which should relieve research assistants who are currently being asked to label images as cat/not-cat.
Dealing with large sets of unstructured data - this is a processing problem for which the current infrastructure is sufficient.
IMO we have plenty of data, we don't need many more labels, and we have sufficient processing power. That leaves algorithmic work for the discovery of causes. (My bet is on systems similar to the kind of sparse distributed hierarchical models that Numenta is working on.)
One thing that I don't think most people have realized is how far past the quantity of data and processing power we need we already are.
If you were to take all the visual information the brain receives in the first 30 years of life (plenty of time to become intelligent) you'd arrive at a number around 1.4 petabytes. This is hardly a large number compared to the amount of information processed by YouTube on a monthly basis. Hearing, taste, and smell all add in a fraction more, and touch (well I don't have any good numbers there yet).
The hard part now is going to be convincing machine learning researchers to actually expose their algorithms to a swath of data similar to the inputs we humans get, over sufficient time, so as to allow for the development of a robust hiearchy of stored causes (aka understanding.)
AFAICT - Experience is understanding.