Are self-training narrow AIs even a going concern yet? DeepQA can update its knowledge base in situ, but must be instructed to do so. Extracting syntactic and semantic information from a corpus is the easy part; figuring out what that corpus should include is still an open problem, requiring significant human curation. I don’t think anyone’s solved the problem of how an AI should evaluate whether to update its knowledge base with a new piece of information or not. In the Watson case, an iterative process would be something like “add new information → re-evaluate on gold standard question set → decide whether to keep new information”, but Watson’s fitness function is tied to that question set. It’s not clear to me how an AI with a domain-specific fitness function would acquire any knowledge unrelated to improving the accuracy of its fitness function—though that says more about the fitness functions that humans have come up with so far than it does about AGI.
It’s certainly the case that an above-human general intelligence could copy the algorithms and models behind a narrow AI, but then, it could just as easily copy the algorithms and models that we use to target missiles. I don’t think the question “is targeting software narrow AI” is a useful one; targeting software is a tool, just as (e.g.) pharmaceutical candidate structure generation software is a tool, and an AGI that can recognize the utility of a tool should be expected to use it if its fitness function selects a course of action that includes that tool. Recognition of utility is still the hard part.
I wouldn’t call Google’s search personalization “self-training” because the user is responsible for adding new data points to his or her own model; it’s the same online algorithm it’s always been, just tailored to billions of individual users rather than a set of billions of users. The set of links that a user has clicked on through Google searches is updated every time the user clicks a new link, and the algorithm uses this to tweak the ordering of presented search results, but AFAIK the algorithm has no way to evaluate whether the model update actually brought the ordering closer to the user’s preferred ordering unless the user tells it so by clicking on one of the results. It could compare the ordering it did present to the ordering it would have presented if some set of data points wasn’t in the model, but then it would have to have some heuristic for which points to drop for cross-validation.
The way I see it, the difference between an online algorithm and a self-training AI is that the latter would not only need such a heuristic—let’s call it “knowledge base evaluation”—it would also need to be able to evaluate the fitness of novel knowledge base evaluation heuristics. (I’m torn as to whether that goalpost should also include “can generate novel KBE heuristics”; I’ll have to think about that a while longer.) Even so, as long as the user dictates which points the algorithm can even consider adding to its KB, the user is acting as a gatekeeper on what knowledge the algorithm can acquire.
The way I see it, the difference between an online algorithm and a self-training AI is that the latter would not only need such a heuristic—let’s call it “knowledge base evaluation”—it would also need to be able to evaluate the fitness of novel knowledge base evaluation heuristics.
On reflection, I’m now contradicting my original statement; the above is a stab toward an algorithmic notion of “self-training” that is orthogonal to how restricted an algorithm’s training input set is, or who is restricting it, or how. Using this half-formed notion, I observe that Google’s ranking algorithm is AFAIK not self-training, and is also subject to a severely restricted input set. I apologize for any confusion.
Are self-training narrow AIs even a going concern yet? DeepQA can update its knowledge base in situ, but must be instructed to do so. Extracting syntactic and semantic information from a corpus is the easy part; figuring out what that corpus should include is still an open problem, requiring significant human curation. I don’t think anyone’s solved the problem of how an AI should evaluate whether to update its knowledge base with a new piece of information or not. In the Watson case, an iterative process would be something like “add new information → re-evaluate on gold standard question set → decide whether to keep new information”, but Watson’s fitness function is tied to that question set. It’s not clear to me how an AI with a domain-specific fitness function would acquire any knowledge unrelated to improving the accuracy of its fitness function—though that says more about the fitness functions that humans have come up with so far than it does about AGI.
It’s certainly the case that an above-human general intelligence could copy the algorithms and models behind a narrow AI, but then, it could just as easily copy the algorithms and models that we use to target missiles. I don’t think the question “is targeting software narrow AI” is a useful one; targeting software is a tool, just as (e.g.) pharmaceutical candidate structure generation software is a tool, and an AGI that can recognize the utility of a tool should be expected to use it if its fitness function selects a course of action that includes that tool. Recognition of utility is still the hard part.
Is what Google does for search results based in part on what you do and don’t do considered self training?
What I mean is that two people don’t see the exact same Google results for some queries if we were both signed into Google, and in some cases even if we both aren’t. Article: http://themetaq.com/articles/reasons-your-google-search-results-are-different-than-mine
An entirely separate question is whether or not Google is a narrow AI, but I figured I should check one thing at a time.
I wouldn’t call Google’s search personalization “self-training” because the user is responsible for adding new data points to his or her own model; it’s the same online algorithm it’s always been, just tailored to billions of individual users rather than a set of billions of users. The set of links that a user has clicked on through Google searches is updated every time the user clicks a new link, and the algorithm uses this to tweak the ordering of presented search results, but AFAIK the algorithm has no way to evaluate whether the model update actually brought the ordering closer to the user’s preferred ordering unless the user tells it so by clicking on one of the results. It could compare the ordering it did present to the ordering it would have presented if some set of data points wasn’t in the model, but then it would have to have some heuristic for which points to drop for cross-validation.
The way I see it, the difference between an online algorithm and a self-training AI is that the latter would not only need such a heuristic—let’s call it “knowledge base evaluation”—it would also need to be able to evaluate the fitness of novel knowledge base evaluation heuristics. (I’m torn as to whether that goalpost should also include “can generate novel KBE heuristics”; I’ll have to think about that a while longer.) Even so, as long as the user dictates which points the algorithm can even consider adding to its KB, the user is acting as a gatekeeper on what knowledge the algorithm can acquire.
On reflection, I’m now contradicting my original statement; the above is a stab toward an algorithmic notion of “self-training” that is orthogonal to how restricted an algorithm’s training input set is, or who is restricting it, or how. Using this half-formed notion, I observe that Google’s ranking algorithm is AFAIK not self-training, and is also subject to a severely restricted input set. I apologize for any confusion.