Machine learning needs stimulus for making solutions.
"A barren plateau is a trainability problem that occurs in machine learning optimization algorithms when the problem-solving space turns flat as the algorithm is run. Researchers at Los Alamos National Laboratory have developed theorems to prove that any given algorithm will avoid a barren plateau as it scales up to run on a quantum computer." (https://www.lanl.gov/discover/news-release-archive/2021/March/0319-barren-plateaus.php)
Have you heard of a "barren plateaus" problem? The name of that problem is coming from the J.R.R Tolkien book. Wherein the fictive Middle land is the very dry and hot place. The image, above this text, introduces the "barren plateaus" problem very well.
In the next example, the food and water are the information. If a creature lives in the "fresh plateaus" it can take nutrients from nature. There is a bigger chance to make mistakes. But the nutrient is versatile and finding and testing new things makes the creature in the work. When a creature searches for things from nature there are many possibilities to test which type of vegetables or other nutrient sources the creature uses. Of course, that thing requires sometimes rise to the mountains.
In that image, the problem is the mountain. The thing is that if the creature lives on the landscape at the higher image. That creature has stimulus. The green landscape offers motivation and the grass is the food and the creature wants to go to the mountain. The grass and water are everywhere and the creature has the motivation to rise to the mountain. And that could be willing to see the farter places. Or maybe the creature wants to get fresh air.
The lower image introduces the situation. Where the creature lives in the "barren plateau". The water and food are in a pocket or bag and of course, the creature never makes mistakes if that creature wants to get a certain sandwich. The creature knows which pocket that creature can get sausage sandwich and where is the drinking bottle. But sooner or later, the nutrient would turn unilateral. In barren plateaus problem, the creature will get pre-made food.
So if we are transferring that model to the information technology the creature will get pre-made solutions that fit in certain situations. And that makes this kind of model very limited. The situation is like that creature lives in the desert or "barren plateaus". The supporter brings water and sandwich to a certain point at a certain time. The food is guaranteed but it's always the same.
And what the creature gets depends on the supporter. If the supporter wants to give the sausage sandwich that is the food. If someday the supporter wants to give the cheese sandwich that creature will get a cheese sandwich.
When everything is pre-made the creature doesn't want to try itself to find food. There is difficult to make mistakes if some other person makes the food. The same way is in data science. If all problems are pre-solved that thing means that it's very hard to make wrong solutions.
The term "flatten landscape" means that when the creature is living in "barren plateaus" the limited information sources makes the problems look harder to solve. Because the creature always is at a certain point where the supporter will bring a sandwich and water the creature is not even trying to climb mountains or solve the problem.
"A barren plateau is a trainability problem that occurs in machine learning optimization algorithms when the problem-solving space turns flat as the algorithm is running".
"In that situation, the algorithm can’t find the downward slope in what appears to be a featureless landscape and there’s no clear path to the energy minimum. Lacking landscape features, machine learning can’t train itself to find the solution". (LosAlamos National laboratories, Solving ‘barren plateaus’ is the key to quantum machine learning)
https://www.lanl.gov/discover/news-release-archive/2021/March/0319-barren-plateaus.php
Image:https://www.lanl.gov/discover/news-release-archive/2021/March/0319-barren-plateaus.php
Comments
Post a Comment