I wonder if we could articulate a Moore's law for machine learning. Every X months, we see a doubling of Y, where Y represents something intrinsic like transistor count.
zdnet.com/article/google…
zdnet.com/article/google…
The obvious kinds of metrics are "papers" or "funding", but those are easily gamed.
Another idea is a battery of cognitive tests, like an AIQ score, but that's also gameable.
Ideally we want a ratio-scale variable like transistor count. Maybe neural network size?
Another idea is a battery of cognitive tests, like an AIQ score, but that's also gameable.
Ideally we want a ratio-scale variable like transistor count. Maybe neural network size?
Maybe we could define an AIQ score after all. Take AI vs human performance on various benchmarks, and plot each of those over time as shown.
I'd refresh a regularly updated dashboard with all benchmarks like this. Maybe rise of the machines dot com?
gigazine.net/gsc_news/en/20…
I'd refresh a regularly updated dashboard with all benchmarks like this. Maybe rise of the machines dot com?
gigazine.net/gsc_news/en/20…
Human performance remains roughly constant in these things, while machine performance improves. So it's a possible benchmark.
Score each as 0.1X as good as a human, 1X as good, 10X as good, and so on.
To aggregate you might say "in 2017, AI was >1X human in 42% of benchmarks."
Score each as 0.1X as good as a human, 1X as good, 10X as good, and so on.
To aggregate you might say "in 2017, AI was >1X human in 42% of benchmarks."
