Timeline for Why not just dump the neural networks and deep learning?
Current License: CC BY-SA 3.0
12 events
when toggle format | what | by | license | comment | |
---|---|---|---|---|---|
Jun 25, 2018 at 12:47 | history | made wiki | Post Made Community Wiki by whuber♦ | ||
Aug 11, 2017 at 15:59 | history | edited | Carrosive | CC BY-SA 3.0 |
deleted 275 characters in body
|
Aug 11, 2017 at 15:55 | comment | added | Carrosive | @klumbard Oh I can see where you are coming from now. So although the model produces estimates that are unlikely to be completely accurate, we can measure the error term to state how much the real values may deviate from the estimates, and thus it would be incorrect to say that the model is inherently wrong. I'll take that part out of my answer, I think my point is better explained in the part I added after it. Thanks for explaining :) | |
Aug 11, 2017 at 15:17 | comment | added | klumbard | The salient point, which you do address, is that all models are wrong because of omitted variable bias as well as misspecification of the functional form. Every time you write down a regression model and perform inference on the estimates, you are assuming you have correctly specified the model, which is never the case. | |
Aug 11, 2017 at 15:16 | comment | added | klumbard | My only issue with your post is the way you phrase "...as in many cases our observations and predictions will not sit exactly on the fitted line. This is one way in which our model is often 'wrong'...". I'm simply saying that the specification of the model includes an error term and so the fact (alone) that the observed data do not fall on the fitted line does not indicate model "wrongness". This might seem like a subtle semantic distinction but I think it's important | |
Aug 11, 2017 at 15:02 | history | edited | Carrosive | CC BY-SA 3.0 |
added 4 characters in body
|
Aug 11, 2017 at 14:57 | comment | added | Carrosive | @klumbard Thanks for the comment. I have updated my answer with more detail which explains my reasoning behind using this as an example. I took a more philosophical approach in my answer and spoke in more general terms rather than specifics, this is my first post in this community so apologies if this is not the place to do so. You seem knowledgeable about the specifics, could you elaborate on your comment a bit more? The question I have is, where deviations do not demonstrate deficiency, is a regression model with an R-squared of 0.01 also not "wrong"? | |
Aug 11, 2017 at 14:34 | history | edited | Carrosive | CC BY-SA 3.0 |
added 1368 characters in body
|
Aug 11, 2017 at 12:48 | comment | added | klumbard | You make some good points, but the fact that "in many cases our observations and predictions will not sit exactly on the fitted line" is not an apt demonstration of the "all models are wrong" slogan. In linear regression we are modelling E(Y|X) and thus points not lying exactly on the line do not demonstrate a deficiency in our model. Randomness is prespecified and expected; the model is not "wrong" when we observe deviations from the fitted line. | |
Aug 11, 2017 at 11:53 | history | edited | Carrosive | CC BY-SA 3.0 |
added 64 characters in body
|
Aug 11, 2017 at 11:50 | review | First posts | |||
Aug 11, 2017 at 12:00 | |||||
Aug 11, 2017 at 11:47 | history | answered | Carrosive | CC BY-SA 3.0 |