I was browsing a discussion when I saw that a lot of question answering systems seek to provide the type of answers that one might gain from doing an internet search or an encylopaedia search.
This is a fine application but I have a vision that more can be done. Question answering systems potentially can do higher levels of reasoning. At present question answering systems can answer questions that seek facts, descriptions about things and simple relationships. A higher level of reasoning would involve having a set of concepts and using those concepts to come to a result. I can only think of very advanced examples that we don't know how to approach at present.
An advanced example might be giving a pathfinding algorithm that is written in natural language to an artificial intelligence and then having that artificial intelligence apply that algorithm.
Another advanced example might be giving a set of natural language instructions on chess strategies to an artificial intelligence and then watching that artificial intelligence apply those strategies.
Another example is to ask an artificial intelligence to to temporarily anwer "3" the question "What is 1+1?." An artificial intelligence of the type I am looking would answer 3. Most artificial intelligence programs such as chat bots would answer 2. Most question answering systems would answer with whatever scrap of information they found in the group of documents they have as reference.
Please don't get me wrong. The systems being developed are great, though I believe they can be so much more and by being so they can become so much more useful and relevant to us. The key is to hold on to the vision of what they can be; to hold on to specific performance functionality that I feel is very possible in the near future.
The choice is ours. We can keep going with the current approach of taking what we know and encoding that knowledge for an artificial intelligence to blindly follow or we can encode the right knowledge for an artificial intelligence to be able to learn more for itself.
The information bottleneck needs to be recognised instead of being worked around. It has been shown time and again that we know how to take human knowledge and encode it for a machine. This is useful though limited. It means that every type of knowledge requires a human to interpret for every single possible situation which is infinite.
The other way would be to as a preliminary step develop a system that has enough initial knowledge and capabilities to learn from the massive store of written natural language documents. This would be an indicator to whether it is possible to enable an artificial intelligence to learn for itself in a more powerful way.
This is a fine application but I have a vision that more can be done. Question answering systems potentially can do higher levels of reasoning. At present question answering systems can answer questions that seek facts, descriptions about things and simple relationships. A higher level of reasoning would involve having a set of concepts and using those concepts to come to a result. I can only think of very advanced examples that we don't know how to approach at present.
An advanced example might be giving a pathfinding algorithm that is written in natural language to an artificial intelligence and then having that artificial intelligence apply that algorithm.
Another advanced example might be giving a set of natural language instructions on chess strategies to an artificial intelligence and then watching that artificial intelligence apply those strategies.
Another example is to ask an artificial intelligence to to temporarily anwer "3" the question "What is 1+1?." An artificial intelligence of the type I am looking would answer 3. Most artificial intelligence programs such as chat bots would answer 2. Most question answering systems would answer with whatever scrap of information they found in the group of documents they have as reference.
Please don't get me wrong. The systems being developed are great, though I believe they can be so much more and by being so they can become so much more useful and relevant to us. The key is to hold on to the vision of what they can be; to hold on to specific performance functionality that I feel is very possible in the near future.
The choice is ours. We can keep going with the current approach of taking what we know and encoding that knowledge for an artificial intelligence to blindly follow or we can encode the right knowledge for an artificial intelligence to be able to learn more for itself.
The information bottleneck needs to be recognised instead of being worked around. It has been shown time and again that we know how to take human knowledge and encode it for a machine. This is useful though limited. It means that every type of knowledge requires a human to interpret for every single possible situation which is infinite.
The other way would be to as a preliminary step develop a system that has enough initial knowledge and capabilities to learn from the massive store of written natural language documents. This would be an indicator to whether it is possible to enable an artificial intelligence to learn for itself in a more powerful way.
Powered by ScribeFire.
No comments:
Post a Comment