Combining learned skills
What helps a model reuse learned skills on unfamiliar problems, and how can that ability be measured?
Research
Research questions
The work starts with questions that are easy to state and hard to answer: Can a model handle something new? Can it stay on task, remember what matters, and learn from what happens?
Areas
What ExoSigma wants to find out
Long tasks
What helps a system keep track of the goal, notice when it is off course, and recover without starting over?
Useful memory
Which kinds of memory make a system more useful over time, and which make it brittle or hard to understand?
Learning during use
When is it useful for a model to update during a task, and how can those changes be evaluated fairly?
Ideas borrowed from neuroscience
Can an idea from neuroscience be turned into a useful computational mechanism, or does a simpler method work just as well?
Evaluation
What would count as evidence?
Every experiment needs a baseline and a result that would count against the idea. A public writeup should include enough detail for someone else to follow the comparison.
- Baseline
- Compare against the strongest relevant simple baseline.
- Test
- Choose tests where competing ideas lead to different outcomes.
- Claims
- Say only what the result supports.
- Record
- Keep enough context for someone else to follow how the result was produced.
No project pages yet.
Project pages will be added when there is work ready to share.
Until then, this page describes questions, not results.