Untold History of AI: Algorithmic Bias Was Born in the 1980s | Hacker News Reader

Untold History of AI: Algorithmic Bias Was Born in the 1980s

spectrum.ieee.org

Tangential comment: When I first learned how long ML has been around I was really surprised. Markov chains were invented more than a hundred years ago. The perceptron was invented in the late 50s - and it's clear now that just a simple perceptron can do a lot of fairly impressive things. Backprop was first worked on in the late 60s. RNNs have been around since the 80s.

It feels like ML is such a recent thing, but it's really just that we've only recently gotten to the levels of hardware and data needed to make these algorithms work smoothly. It makes me wonder what other interesting theories are still gated by hardware limitations. I think astronomy has had similar situations where things have been figured out long before they were actually observed.

Machine learning libraries have been around for a long time also. I remember seeing them around and playing with them in the late 90s, without really knowing what to do with them. For some reason, few people cared about them then, much like few people took virtual reality very seriously at the time. Interest in these things swings back and forth like a pendulum. It seems that you can pick almost any uninteresting thing that no one cares about anymore and work on it, and soon enough, it will be back in the spotlight again.

At one point, naive Bayesian classifiers were considered machine learning. Now we use them for spam filters. Chess playing algorithms were also considered machine learning at one point. We keep moving the goalposts. We won't ever reach a specific point of concluding that we have intelligent machines. There won't be a breakthrough, much like living organisms never had an intelligence breakthrough; they just gradually became more and more intelligent through incremental development.

In the 1980s, they had to explicate their biases in order to make the algorithm's predictions correlate well with human judgment. Today we can just use neural nets to learn our biases for us.

Reply on news.ycombinator.com