Comment by jononor
8 years ago
To use machine learning parlance, such a solution would (likely) be overfitting the problem, and not generalize well. If one instead changed the setup to be: 1) Specify a handful of tasks for the AI system to complete 2) Test the performance on a _separate_ (un)related set of tasks
The test set has to be unknown to the system developers.
If the system can realize the unknown tasks without further input from researchers, in the same way that a mouse can, then we have some level of generalizable intelligence.
What is the baseline? How well does a mouse perform when placed in an "unrelated" task for the first time? The mouse also gets an explicit reward function (food, pain, etc) - does the "unrelated" task use the same reward function as what the AI was optimized for?
Also, is it ever really the "first time" for a mouse when behavior has been ingrained and tuned over millions of years of evolution? Is this different than training an algorithm?
My point is just that it's really hard to define these tasks and how to evaluate performance for a machine and a mouse.
The baseline performance would be whatever a set of mice would do in the same situation. What constitutes an "unrelated" task is quite a difficult question, we would probably need to iterate a lot on that. If we are to have "hidden" tasks available we need to come up with a lot of new task formulations/variations anyway.
I think that replicating mouse-level adaptability in an intelligent agent while allowing 'inherited' behavioral traits will already be an achievement. And probably take us quite a while.