The more complex your model, and the more complex reality is, the closer the correspondence between them, and the more your internal model acts as if it is "learning something" (making incorrect predictions, processing the data, then making better ones), the less scope these is for your symbols to be ungrounded.
That seems to merely assert what I was arguing against .. I was arguing that preditictve accuracy is orthogonal to ontological correctness...and that grounding is to do with ontological correctness.
It's always possible, but the level of coincidence needed to have the wrong model that behave exactly the same as the right one is huge.
Right and wrong don't have univocal meaning, here. A random model will have poor predictive accuracy, but you can still have two models of equivalent predictive accuracy, but different ontological implications.
And, I'd say, having the wrong model that gives the right predictions is just the same as having the randomized labels.
You seem to be picturing a model as a graph with labelled vertices, and assuming that two equally good models most have the same structure. That is not so.
For instance, the Ptolemaic system can be made as accurate as you want for generating predictions, by adding extra epicycles ... although it is false, in the sense of lacking ontological accuracy, since epicycles don't exist.
Another way is to notice that ontological revolutions can make merely modest changes to predictive abilities. Relativity inverted the absolute space and time of Newtonian physics, but its predictions were so close that subtle experiments were required to distinguish the two,,
In that case, there is still, a difference in empirical predictiveness. In the extreme case there is not: you can have two ontologies that always make the sane predictions, the one being dual to the other. An example is wave particle duality in quantum mechanics.
The fourth way is based on sceptical hypotheses, such as Brain in a Vat and Simulated Reality. Sceptical hypotheses can be rejected, for instance by appeals to Occams Razor, but they cannot be refuted empirically, since any piece of empirical evidence is subject to sceptical interpretation. Occams's Razor is not empirical.
Science conceives of perception as based in causation, and causation as being comprised of chains of causes and effects, with only the ultimate effect, the sensation evoked in the observer, being directly accessible to the observer. The cause of the sensation, the other end of the causal chain, the thing observed, has to be inferred from the sensation, the ultimate effect -- and it cannot be inferred uniquely, since, in general, more than one cause can produce the same effect. A further proxy can always be inserted into a series of proxies. All illusions, from holograms to stage conjuring, work by producing the effect, the percept, in an unexpected way. A BIV or Matrix observer would assume that the precept of a horse is caused by a horse, but it would actually by a mad scientist pressing buttons.
A BIV or Matrix inhabitant could come up with science that works, that is useful, for many purposes, so long as their virtual reality had some stable rules. They could infer that dropping an (apparent) brick onto their (apparent) foot would cause pain, and so on. It would be like the player of a computer game being skilled in the game, knowing its internal physics.The science of the Matrix inhabitants would work, in a sense, but the workability of their science would be limited to relating apparent causes to apparent effects, not to grounding causes and effects in ultimate reality. But empiricism cannot tell us that we are not in the same situation.
In the words of Werner Heisenberg (Physics and Philosophy, 1958) "We have to remember that what we observe is not nature herself, but nature exposed to our method of questioning"
We don't seem to be disagreeing about anything factual. You just want grounding to be in "the fundamental ontology", while I'm content with them being grounded in the set of everything we could observe. If you like, I'm using Occam or simplicity priors on ontologies; if there are real objects behind the ones we can observe but we never know about them, I'd still count our symbols as grounded. (that's why I'd count virtual Napoleon's symbols as being grounded in virtual Waterloo, incidentally)
Myself, Kaj Sotala and Seán ÓhÉigeartaigh recently submitted a paper entitled "The errors, insights and lessons of famous AI predictions and what they mean for the future" to the conference proceedings of the AGI12/AGI Impacts Winter Intelligence conference. Sharp deadlines prevented us from following the ideal procedure of first presenting it here and getting feedback; instead, we'll present it here after the fact.
The prediction classification shemas can be found in the first case study.
Locked up in Searle's Chinese room
Searle's Chinese room thought experiment is a famous critique of some of the assumptions of 'strong AI' (which Searle defines as the belief that 'the appropriately programmed computer literally has cognitive states). There has been a lot of further discussion on the subject (see for instance (Sea90,Har01)), but, as in previous case studies, this section will focus exclusively on his original 1980 publication (Sea80).
In the key thought experiment, Searle imagined that AI research had progressed to the point where a computer program had been created that could demonstrate the same input-output performance as a human - for instance, it could pass the Turing test. Nevertheless, Searle argued, this program would not demonstrate true understanding. He supposed that the program's inputs and outputs were in Chinese, a language Searle couldn't understand. Instead of a standard computer program, the required instructions were given on paper, and Searle himself was locked in a room somewhere, slavishly following the instructions and therefore causing the same input-output behaviour as the AI. Since it was functionally equivalent to the AI, the setup should, from the 'strong AI' perspective, demonstrate understanding if and only if the AI did. Searle then argued that there would be no understanding at all: he himself couldn't understand Chinese, and there was no-one else in the room to understand it either.
The whole argument depends on strong appeals to intuition (indeed D. Dennet went as far as accusing it of being an 'intuition pump' (Den91)). The required assumptions are:
Thus the Chinese room argument is unconvincing to those that don't share Searle's intuitions. It cannot be accepted solely on Searle's philosophical expertise, as other philosophers disagree (Den91,Rey86). On top of this, Searle is very clear that his thought experiment doesn't put any limits on the performance of AIs (he argues that even a computer with all the behaviours of a human being would not demonstrate true understanding). Hence the Chinese room seems to be useless for AI predictions. Can useful prediction nevertheless be extracted from it?
These need not come directly from the main thought experiment, but from some of the intuitions and arguments surrounding it. Searle's paper presents several interesting arguments, and it is interesting to note that many of them are disconnected from his main point. For instance, errors made in 1980 AI research should be irrelevant to the Chinese Room - a pure thought experiment. Yet Searle argues about these errors, and there is at least an intuitive if not a logical connection to his main point. There are actually several different arguments in Searle's paper, not clearly divided from each other, and likely to be rejected or embraced depending on the degree of overlap with Searle's intuitions. This may explain why many philosophers have found Searle's paper so complex to grapple with.
One feature Searle highlights is the syntactic-semantic gap. If he is correct, and such a gap exists, this demonstrates the possibility of further philosophical progress in the area (in the opinion of one of the authors, the gap can be explained by positing that humans are purely syntactic beings, but that have been selected by evolution such that human mental symbols correspond with real world objects and concepts - one possible explanation among very many). For instance, Searle directly criticises McCarthy's contention that ''Machines as simple as thermostats can have beliefs'' (McC79). If one accepted Searle's intuition there, one could then ask whether more complicated machines could have beliefs, and what attributes they would need. These should be attributes that it would be useful to have in an AI. Thus progress in 'understanding understanding' would likely make it easier to go about designing AI - but only if Searle's intuition is correct that AI designers do not currently grasp these concepts.
That can be expanded into a more general point. In Searle's time, the dominant AI paradigm was GOFAI (Good Old-Fashioned Artificial Intelligence (Hau85)), which focused on logic and symbolic manipulation. Many of these symbols had suggestive labels: SHRDLU, for instance, had a vocabulary that included 'red', 'block', 'big' and 'pick up' (Win71). Searle's argument can be read, in part, as a claim that these suggestive labels did not in themselves impart true understanding of the concepts involved - SHRDLU could parse ''pick up a big red block'' and respond with an action that seems appropriate, but could not understand those concepts in a more general environment. The decline of GOFAI since the 1980's cannot be claimed as vindication of Searle's approach, but it at least backs up his intuition that these early AI designers were missing something.
Another falsifiable prediction can be extracted, not from the article but from the intuitions supporting it. If formal machines do not demonstrate understanding, but brains (or brain-like structures) do, this would lead to certain scenario predictions. Suppose two teams were competing to complete an AI that will pass the Turing test. One team was using standard programming techniques on computer, the other were building it out of brain (or brain-like) components. Apart from this, there is no reason to prefer one team over the other.
According to Searle's intuition, any AI made by the first project will not demonstrate true understanding, while those of the second project may. Adding the reasonable assumption that it is harder to simulate understanding if one doesn't actually possess it, one is lead to the prediction that the second team is more likely to succeed.
Thus there are three predictions that can be extracted from the Chinese room paper:
Therefore one can often extract predictions from even the most explicitly anti-predictive philosophy of AI paper.
References: