patternMinor
Why do black-box deep learning algorithms produce better word similarities than cognitively plausible ones?
Viewed 0 times
whyproducesimilaritiesthanbetteralgorithmslearningcognitivelydeepword
Problem
Why do black-box deep learning algorithms (e.g. Mikolov et al 2013) produce better word similarities than cognitively plausible ones (e.g. Baayen et al 2011)? Could we combine the best of both worlds, biologically motivated learning mechanisms and state-of-the-art performance?
Solution
The short answer is that we don't know why current ML algorithms work well on a particular dataset, or why other ML algorithms don't.
Most researchers focus on finding ML algorithms that achieve high accuracy, regardless of how they achieve it. Cognitive plausibility typically isn't a goal.
Generally, across much of computer science, our best methods often differ from how the brain works. This is true for machine learning, for computer vision, for playing games like chess, and other tasks. That could be viewed as an opportunity for improvement and a failure of our methods, or as an indication that the methods that will work best for computers might be different from the ones that our brains have evolved. We just don't know.
Most researchers focus on finding ML algorithms that achieve high accuracy, regardless of how they achieve it. Cognitive plausibility typically isn't a goal.
Generally, across much of computer science, our best methods often differ from how the brain works. This is true for machine learning, for computer vision, for playing games like chess, and other tasks. That could be viewed as an opportunity for improvement and a failure of our methods, or as an indication that the methods that will work best for computers might be different from the ones that our brains have evolved. We just don't know.
Context
StackExchange Computer Science Q#53854, answer score: 2
Revisions (0)
No revisions yet.