Machine learning and AI which are fancy names of the old operations research procedures of the sixties for minimization such a the Gradient methods, usually based now a days on neural networks, neuro-fuzzy systems , genetic algorithms e.t.c .
The question hence arises: How do we know that AI and machine learning based minimization procedures do not result to a local minimum??
This should answer your question:
Article Local minima in training of deep networks
https://stats.stackexchange.com/questions/203288/understanding-almost-all-local-minimum-have-very-similar-function-value-to-the
Dear Peter,
In fact the preliminary analysis of the data can avoid this problem, it is possible to use as an example a neural network of learning (or other technique belonging to the AI) and to make several tests on this last one to see which points of initialization and what data from the space in question can lead to situations of total convergence.
Also you can take a look at links in subject.
-45 Questions to test a data scientist on Deep Learning (along with ...
https://www.analyticsvidhya.com › Machine Learning
-Lecture 4: Machine learning III Question Review Review
https://web.stanford.edu/class/cs221/lectures/learning3-6pp.pdf
Best regards
If the target function is convex, local optimal point is global optimal point. Otherwise, we can not assure that whether a result is global optimal point. AI algorithms such as neural network and social behavior algorithms (PSO, ABC) give out approximations or expectations of global optimization. However, in theory, there is many researches relevant to global optimization with non-convex function. Please read the book "Convex Analysis and Global Optimization" by Hoang Tuy. I also proposed an algorithm for searching global minimum point with arbitrary target function in the article "Article Global Optimization with Descending Region Algorithm
".
Your question is absolute nonsense. You may have a grudge against machine learning researchers and despise them for some personal reasons, but there is no point in carrying that to Reseach Gate. If you were really familiar with machine learning, you would know that there is much, much more to machine learning than just optimization. All serious researchers in machine learning are perfectly aware, and acknowledge, that optimization methods used in our field were invented long before machine learning came to existence; but optimization is, by far, not the most important issue in machine learning.
Furthermore, the problem of local minima has been extensively investigated, for instance in relation to neural network training. The concern is not to reach the global minimum of a cost function: it is to find the model that generalizes best given the available data. In most cases, the model that generalizes best given the available data is not the model for which the cost function used for training is at its global minimum. Some regularization methods (e.g. early stopping) are specially designed to avoid the global minimum. All this is textbook stuff, anf I suggest that you read a couple of good textbooks on machine learning.
Mutiple minima points can be identified
Local minimum point can be only one of them.Out of the balance the global minimum point can identified as the least valued minimum point
To give an answer as Gerard Dreyfus that the question is an nonsense and that there exists a grudge against the researchers who work on Machine learning is equivalent to give an answer with a negative sign which is not what Research Gate is all about. The other question comes up. I can mention many examples and another which comes to mind quickly is the following Moreover the answer of Dreyfus is contradictory because my question referred to an optimization as a general concept. Even the word "learning" includes in itself the concept of getting continuously better toward the best which is an optimization process.Think how a person learns a foreign language. Dreyfus's phrase
- Badges
- Science topic