AI and Brittlenss
- A lot of AI systems work really well for their particular
domain.
- Those domains are always restricted (simple task, simple domain, or
both).
- When the system expands or is expanded beyond that domain,
it breaks down.
- This is called brittleness.
- People aren't brittle.
- Learning is one way to address this problem.
- Current machine learning is generally restricted to preprocessed data, or
simplified data.
- Learning is certainly part of the solution.
- Indeed, AI researchers frequently propose magic bullets to
resolve this problem.
- For instance, Minsky says it better search mechanisms will resolve
it.
- How can we get around this brittleness?
