AI News: Bad Links Resulted in My Finding This Smart Software Post
January 17, 2016
I followed some bad links and after a bit of poking around, I arrived at “A Tour of Machine Learning Algorithms.” The write up is more than two years old, but it does have some useful information to feed the hungry minds of the smart software crowd.
The write up explains machine learning algorithms group by learning style and by similarity. The brief explanations are helpful.
The major omission is that none of the algorithms apparently has any notable flaws or vulnerabilities. In our research, we found that the algorithms we tested could be manipulated by creating data which threw the method off the scent for accuracy.
One example is the flaw of a properly “trained” Bayesian system. By loading content outside the original training set’s scope, the Bayesian system happily generated off point results. This means that directed content injection such as shaped or weaponized blog posts or social media message flooding can make the systems wander.
Good news for those who want to manipulate smart software. Bad news for the decision makers who “assume’ the Bayesian system’s outputs remain on the mark.
Useful to know in my opinion. Descriptions of algorithms have to include their less sunny side, don’t you think?
Stephen E Arnold, January 17, 2016
Comments
-
- Subscribe to Beyond Search
Feature archive
News archive
-
