Working Up

by Dwayne Phillips

We have yet another problem with machine learning. This one, however, has an easy solution.

Several years ago I wrote of a fundamental problem with machine learning. I guess we have yet another problem being called “poisoned” machine learning.

With poisoning, someone knows that I am about to “learn the machine” something using a large dataset. That poisoner knows where the dataset is and they put data into my dataset. Their data will mislead my “learned machine” into making mistakes like the poisoner wants it to make.

Alas, garbage in and garbage out, again.

Solutions? Why in the world is my dataset out there where others can poison it? Don’t I know anything about data? Can’t I keep my mouth shut about what I am doing? What is wrong with me?

Solutions? Don’t I know what data I am using to “learn my machine?” Why am I using bad data? What is wrong with me?

I guess this is all more complicated than I understand. Otherwise, the solutions wouldn’t seem so obvious to me. Perhaps I am just too old to “get” all this.