Machine Learning Foibles: Are We Surprised? Nope
March 18, 2020
Eurekalert published “Study Shows Widely Used Machine Learning Methods Don’t Work As Claimed.” Imagine that? The article states:
Researchers demonstrated the mathematical impossibility of representing social networks and other co0mplex networks using popular methods of low dimensional embeddings.
To put the allegations and maybe mathematical proof in context, there are many machine learning methods and even more magical thresholds the data whiz kids fiddle to generate acceptable outputs. The idea is that as long as the outputs are “good enough”, the training method is okay to use. Statistics is just math with some good old fashioned “thumb on the scale” opportunities.
The article states:
The study evaluated techniques known as “low-dimensional embeddings,” which are commonly used as input to machine learning models. This is an active area of research, with new embedding methods being developed at a rapid pace. But Seshadhri and his coauthors say all these methods share the same shortcomings.
What are the shortcomings?
Seshadhri and his coauthors demonstrated mathematically that significant structural aspects of complex networks are lost in this embedding process. They also confirmed this result by empirically by testing various embedding techniques on different kinds of complex networks.
The method discards or ignores information, relying on a fuzz ball which puts an individual into a “geometric representation.” Individuals’ social connections are lost in the fuzzification procedures.
Big deal. Sort of. The paper opens the door to many graduate students’ beavering away on the “accuracy” of machine learning procedures.
Stephen E Arnold, March 18, 2020
Comments
-
- Subscribe to Beyond Search
Feature archive
News archive
-
