An article in the department of Reverse Engineering (bka Neurobiology), suggests that a protein called Rac, is partially responsible for "forgetting." Specifically, flies with Rac production suppressed appeared to forget a negative stimulus response more quickly, and vice versa.
Maybe Eternal Sunshine of the Spotless Mind isn't so far off?
It's the versa that's more interesting. Imagine a world where everyone has eidetic memories?
One thing that's relevant to my research is the benefit of forgetting. In ML, there's a pretty consistent theme that we always want more data. This is because ML is a descendant of stats, which is obsessed with asymptotic convergence to the truth. A common problem in stats/ML is "how many samples do I need to be confident with my conclusions?" (usually alot). But ML, as an engineering discipline, is really concerned about the limits of what we can do right now, with finite samples and with finite resources. So forgetting, in a sense, may be optimal with respect to the limits of computation.
I am pretty convinced that nonparametric approaches (Bayesian or otherwise) are the best way in which to get the most accurate models, but in lieu of infinite resources, I am usually faced with the problem of deciding which examples to learn from. Clustering or sampling approaches are usually used (with the attendant icky feeling), or some semi-parametric approaches commonly posed as prior probability distributions (little less icky).
In an online model, deciding which examples to forget is the other side of the coin, and I've seen a few papers on it, but I haven't read any. I'll read some today, I'm really glad I came across the New Scientist. It's my new favorite site.
No comments:
Post a Comment