Are current AI models entirely empiricist? Can they be rationalist?If so How?
It's important to note that a decision making agent can be perfectly rational and still be "wrong" (i.e. make sub-optimal decisions.)
This is primarily because rationality is bounded for all non-trivial problems. The quality of decisions depends on quality of information, time and space available to make the decision in relation to the problem size, and the quality of the analytic processes.
I would say that the current, strong AIs are empiricists, in that the methods are statistical, and based on testing and evaluation of results. For instance, a Monte Carlo Tree Search (MCTS) game algorithm will play billions of games against itself and analyze the results to make the strongest choice for any given state.
But, if one considers any decision-making agent, no matter how primitive or weak, to be a form of AI, then there can be non-empirical AI in the sense that the processes heuristic as opposed to statistical.
MY OPINION ON WHETHER IT'S POSSIBLE TO BUILD ARTIFICIAL INTELLIGENCE ON THE PRINCIPLES OF RATIONALISM
For a long time philosopher's had struggled to explain how the human mind works and two theories had emerged: empiricism and rationalism
Rationalism states human beings genetically have information about the world that enables them to comprehend the world and that we can build on our knowledge using logical reasoning.For example, according to rationalists 1+1=2 seems so obvious to humans simply because we were born with an intuitive idea of it. Empiricism states on the other hand that humans are born with very little or no information and learn only by the power of observation and practice.
Personally, when it comes to the human mind I think that no theory is fully correct. A mix of both is required to explain the miracle that is the human mind.
Coming to computers, It is a known fact that computers and AI models are so far generally based on the empiricistic model of learning. All supervised learning algorithms are empiricist models.
Then there are clustering models.K-means clustering algorithm subdivides data points of a dataset into clusters based on nearest mean values. This behaviour seems rationalist since no prior examples are provided to the system but that is an incorrect perception. Optimal clustering is achieved after thousands of iterations and humans don't require a fraction of that effort to make satisfactory clusters. This interesting example favours rationalism in humans but certainly lowers the possibility of a rationalist AI model.
So can an AI model ever be based on the principles of rationalism?? I believe that to qualify this criteria the computer system running the AI model won't just be a regular computer. IN FACT I have an experiment in mind to test this... lets train a currently empiricist model on certain data (Say a computer vision problem) (After all , the first half-human-half-monkey wasn't born all that smart was he??) I train this system say A for a few years,and randomly take some of its data (maybe feature matrices,etc). Now I take a newer system B. I give it some data from A(i.e initialize some of its feature matrices to that of A instead of 0,0,,0..) and train B. I do this repeatedly on many systems and finally take another system say Z to which I share the previous generation's data but I do not train the system this time.I directly test it, If it does much better than A (before A was trained). I think in some simplified ways it has qualified to be a rationalist AI model.
Conclusion: Unlike most scientists, I think that AI models CAN be RATIONALIST in the future. I also believe that they, in their own way, may even have a sense of right and wrong. However i also believe that all their "intuitive information" can be traced back to an "empiricist observation based model" which passed on what was learnt over generations (exactly how it has occured in human beings, according to me).
An AI model can be based on the principles of rationalism but not necessarily so. This is an open question the answer of which may be both. There are three hypotheses that cover the answer space.
- Intelligence is rational and theorem proving, construction of mathematical models, probabilistic belief models, and rules inference will be the core of future cognitive AI even though in human brains realize these abilities through neural networks.
- Intelligence is behavioral and rarely rational, although disciplines have developed through a scientific culture to attempt to order the world through mathematics, humans and other animals with brains evolve and learn intelligent behavior through successive attempts at gaining and edge.
- Intelligence is neither based on rationalism nor emergent network behavior and the intelligent features of the human brain are merely a reflection of intelligence in the cosmic, physical, and biological realities that have existed in the universe long before the development of brains.
None of the three have been proven and none have been negated. There may be overlaps or gaps between them, but only minor ones.
Current AI models have to be partly empiricist because theory is always developed to explain sets of observations. They must be partially theoretical (but not necessarily rationalist) because the models must be programmed in declarative or algorithmic form.
Rational thought and rationalism are not the same. Rationalism is a philosophic system that involves determinism, causality, and doubt about purpose and meaning. Rational thought is a type of thought that applies rules of logic and inference to the analysis of data and challenges in control and comprehension. The idea that rational thought always leads to rationalism has counter-examples.
- Bob is a staunch rationalist but loves his mother without reason and refuses to recant either his rationalistic worldview or his love for his mother even though the two positions are in rational conflict.
- Science states about itself that it cannot draw a conclusion without making observations, it is self-evident that existence does not depend on observation, and therefore science cannot prove the non-existence of one or more things that may exist and thus reality may extend outside the scope of scientific inquiry.
- That biology emerged contrary to the second law of thermodynamics is proposed to be a matter of luck, which is fantastic to the extreme.
- Liar's Paradox: Epimenides (VI century BC) was a Cretan angry with his fellow-citizens who suggested, "All Cretans are liars." Is this statement true or false? Gödel's incompleteness theorems shocked the world of rationalism by proving that not all true things can be proven.
Returning to AI, we can develop systems that conform to rational thought, but they may not be rationalistic. We can develop systems in a vacuum, without any data or sensory input, but it may not be applicable for any practical task. Some have suggested that mathematics can and to some degree has been mostly developed without data or sensory input and is applicable, which favors the third hypothesis above, that intelligence predates human thought.
In training artificial networks, we know that the drawing of samples for training can't deviate far from the distribution of input when the trained system is deployed and be expected to work, which is strictly empirical.
The question is a good one because the disparity between these last two paragraphs is large, and that gap might make the thirty year predictions to human like thinking computers may again be overly optimistic by orders of magnitude. Maybe not. It's certainly an area of debate and curiosity.