The Singularity

« first day (926 days earlier) ← previous day next day → last day (2197 days later) »

04:07

@DukeZhou https://chat.stackexchange.com/messages/49033246/history makes sense, being garbled when you are stressed.

I know many questions lack basic formatting but succeed through the power of their question itself, however I went though extra work to do bold formatting so users know they only need to read one paragraph: https://ai.stackexchange.com/posts/10539/revisions.

3 hours later…

prosody-Gab Vereable Context

07:02

Actually by "bold" I mean I summarized in a paragraph and then did formatting. Also, I just corrected the title to say "FB Safety Check AI" to reflect fb.com/zuck's exact words from the already-written Mark Zuckerberg "Safety Check is AI" quote.

I did not know I would need to explain to an AI forum that @zuck talks about AI.

AI Meta Questions

07:50

Q: Is FB.com's "Safety Check/"Crisis Response" for AI.stack or WebApp.stack if @zuck called it Artificial Intelligence?

I asked ai.stackexchange.com/questions/10539/should-facebook-safety-check-work-if-an-account-is-stuck-at-a-name-change-checkp after webapps.stackexchange.com/questions/103217/will-facebook-safety-check-work-if-my-account-is-stuck-at-a-checkpoint-page and each time the reason is that question does...

4 hours later…

Rory Alsop

11:55

@prosody-GabVereableContext you haven't made it clear at all. I still honestly cannot understand your original posts nor the meta one

nbro

@DennisSoemers Hi Dennis! If you're willing to have a look at it, I gave an answer to a RL question on Stats, but I'm not completely sure about all the details, especially the last point (about classifying value and policy iterations as bottom-up or top-down algorithms): https://stats.stackexchange.com/q/258907/82135.

1 hour later…

Dennis Soemers

13:05

@nbro Hmmm I think they would be bottom-up dynamic programming actually... but typically not just a single "run" of bottom-up DP, but many consecutive iterations of a complete bottom-up DP run.

Top-down DP would run recursive calls if a required solution to a subproblem were not yet "memoized". We don't do that in Value Iteration / Policy Iteration. We just assume that whatever solution to a subproblem there is (in the beginning often all-zeros initialization) is valid and directly use that (but then by iteratively repeating the entire thing gradually improve things).

This stuff is not necessarily my area of expertise within RL though. Value Iteration / Policy Iteration are nice for theory, but quite boring in practice in my opinion... so don't quote me on what I say about them :D

nbro

13:24

@DennisSoemers Indeed, we do not perform recursive calls, but, in the case of value and policy iteration, can't we consider that the (approximate) solutions are already "memoized"? Anyway, as I described in the answer, I would consider the smallest sub-problem the one where you need to figure out the value of the goal state. So, in that sense, value and policy iterations are not necessarily bottom-up, because they don't necessarily start from the goal state.

Dennis Soemers

13:54

In VI/PI, WITHIN A SINGLE ITERATION, we essentially treat every single entry in the "table" as an already-solved "smallest sub-problem" (and indeed the order in which we choose to update table entries is generally a free choice, the algorithm works regardless of the order in which we choose to loop, although certain orders may be more efficient than others)

5 hours later…

DuttaA

18:56

@NeilSlater @ hey what do you think about the answers on this question: ai.stackexchange.com/questions/10431/… . IMO the answers are quite wrong to equate loss with accuracy. Or am I wrong?

Neil Slater

19:17

@DuttaA The question is a bit vague and answers not great quality. I upvoted ssegvic's answer because it identifies to look a little deeper, and is semi-correct about relationship between loss and accuracy.

What OP should do is have a CV and maybe also hold-out test set (it is not at all clear that they are doing this) and should pick the metric for model selection before deciding between models. Additional metrics and investigations might be useful to help get better results, but the metric should be a proxy for some desirable outcome

This is often harder to see/understand when people are learning ML and trying to pick "best" models without context. But in the real world, the model is going to be used for something. The metric should measure its performance at that, as closely as possible

Whilst a Loss or Cost function is a necessary construct to get the ML to work at all. It is fine to use one as the deciding metric.

Another possibility for the OP here is that 80% accuracy is trivial (80% of items are in one class, so guessing that all of them are gives you 80% accuracy). If the OP is reporting loss on the training set, then it looks a lot like they have a problem with overfitting.