Jethro's Braindump

How To Know - Celeste Kidd

Machine Learning, NeurIPS
Celeste Kidd | How to Know · SlidesLive

Why are we curious about some things but not others? What are the core cognitive systems that people use to guide their learning about the world?

5 Key things in ML that we should know about humans

Humans continuously form beliefs

Upon seeing a bowl, we update our belief about what a bowl is. This also influences what we wish to sample next. Infants look away when events are predictable, but also when they are surprising!

When present data to learning algorithms, we should show data that is at ideal region on the “surprising” spectrum.

Certainty diminishes interest

We are not curious about things that we already know. If we think we know an answer, when presented with the objective truth, we are still unlikely to change our beliefs.

Certainty is driven by feedback

When feedback isn’t available, our estimates about our certainty may not be accurate. When feedback is readily available, we can sometimes be certain, when we should not be.

Less feedback may encourage overconfidence

Humans form beliefs quickly