I understand that IA is based on complex neural network. I have none experience with AI, but have already used neural networks to build black-box models trying to substitute phenomenological models. I had learned that neural networks work reasonably well interpolating data, but their performance can be poor if one tries to extrapolate from the data domain.I'm horrified to think of asking an AI how to solve climate effects knowing that greenhouse gases come mostly from human activity.
Any comments?
You should be. AI without any Moral controls would reach the obvious logical conclusion that since people are the problem, you can fix it by getting rid of most of the people. This demonstrates the classic flaw of not stating the problem correctly because of prior assumptions. The goal should not be to validate the implied belief that climate effects can be mitigated by eliminating (or at least reducing) greenhouse gases in the atmosphere, but rather generating solutions to the more basic problem of global warming. [To me it is the difference of asking for a selection from a set of predefined solutions vs generation of a new one.] This generation is what I call real AI, and why the selection of the "training" data is critical to eliminate implicit biasing. If you want to get a better perspective, I would recommend going back to the basics. View the historical records of Geoff Hinton's struggles to get AI accepted (see https://www.technologyreview.com/2020/11/03/1011616/ai-godfather-geoffrey-hinton-deep-learning-will-do-everything/ )and the original work at MIT on the Shakey project (the first recorded computer invention of a solution not preprogrammed see https://www.forbes.com/sites/gilpress/2020/01/20/12-ai-milestones-shakey-the-robot/?sh=126f9c585400).
The problem is that current machine learning is being assumed to solve everything and it does not.
The first thing to note is that all machine learning starts with observations from collected data. Even assuming that this process is perfect the machine learning algorithm will then try to model an empirical distribution based on the dataset. The assumption is that based on the data the probability distribution incorporates a priori knowledge based on the regularity of the data from which it was trained.
The problem with the assumption is that it is assumed that the data is "ideal" and that it is enough to extract a regularity from it. The next assumption is that the regularity that is inferred from the data is the desired one (most data can have multiple regularities that can be extracted).
Now to your last comment:
"I'm horrified to think of asking an AI how to solve climate effects knowing that greenhouse gases come mostly from human activity".
There are serious flaws in even asking this question if you think about it based on the workings of machine learning. The reasons are as follow:
machine learning is useful as a tool but in its current form it does not substitute actual thinking.
You are correct that artificial intelligence (AI), including neural networks, can be prone to errors when extrapolating beyond the range of the data used to train them. This is known as the "generalization problem" in AI. To prevent AI from extrapolating and making errors, it is important to carefully select the data used for training and testing, and to use appropriate techniques such as regularization and cross-validation to evaluate the performance of the model. Additionally, it is important to understand the limitations of AI and to use it as a tool to supplement, rather than replace, human expertise and judgment. In the case of climate change, AI can be a valuable tool for analyzing data and making predictions, but it should be used in conjunction with other approaches and with careful consideration of the potential biases and limitations of the data and models used.
- Badges
- Science topic
- Similar topics
- Psychology