Another anecdotal case of different ways that AI learns was the failure in DoD project, when the system had to take millions of photos made by American spy satellites and find the tanks on those photos. In the final analysis the system was ably to pinpoint tanks on many photos but NOT SEEING the obvious tanks on others. After a long study of human and computer psychologists of the cause of that failure they found out that the system was taught to understand that there is a tank by using photos done during sunny days. In some forms of AI it is possible to check how exactly the learning goes, what is the logic, etc. But in some forms of AI (like neural networks) nobody knows how the system learns and thinks. You just have to TRUST it. So it turned out that in the case of tank recognition system devised its own way (somewhat inhumane) in deciding if there is a tank there – by its shadow! Then if on a gloomy day there was no shadow – there was no tank.
logo
logo
Monday, August 1, 2011
More on AI
Speaking of neural networks... You all heard that the machine is only as smart as the smarts that man puts into it. Well, this is not true. A huge part of modern AI is dedicated to machine learning, when AI systems can learn and become MUCH smarter than all programmers put together. But... they learn differently and their “smarts” might be different from human. For example,the system charged with keeping water in NYC clean might start adding poison to the drinking water after concluding that the problem of dirty water is in human activity. It means that a bunch of other thoughts and moral judgments should be included in the systems that can potentially harm people.
Subscribe to:
Post Comments (Atom)
I understand the concept of AI teaching itself new information based on the old information. I don't understand the examples. Can you please give a simpler one?
ReplyDeleteAnother example is face recognition when the system is being shown several photos of the same person and explaining that it is, say, John Doe. Then same teaching to recognize another person. After that the system somehow learns to distinguish these people (nobody might really know how it does it) but after showing totally new photos of different people it will be able to recognize the ones shown before (although that could be different photos, not exactly the ones shown first).
ReplyDeleteAI used in data mining systems can look at the transactions showing who bought what item and figure out what types of products is better to market to what groups of buyers (including the products and the types of buyers that were not in the database). This shows ability to generalize from learned information to something never seen before.
For something like vacuum cleaner RUMBA – AI can plan for some experiments on room exploration from which it learns what areas have to be vacuumed and what is the optimal sequence of actions to clean the whole room/house. Note that nobody taught it how to do it and AI planned own learning, when it was sufficient to stop it, and how to devise optimal plans for required operations.
In addition to gaining intelligence from learning, some forms of AI can devise new knowledge by reasoning about the known facts (just like thinking and coming to realization about some new features of the situation that were not taught it by the designer. In this case it is possible to ask the system produce the chain of reasoning and check if it is correct or adequate.