Timeline for Why do parameters go untested in Machine Learning?
Current License: CC BY-SA 4.0
12 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Apr 14, 2021 at 11:02 | comment | added | Richard Hardy | @DemetriPananos, thank you, I have now learned something new. | |
| Apr 13, 2021 at 16:30 | comment | added | Demetri Pananos | @RichardHardy It is a hyperbaton | |
| Apr 13, 2021 at 16:11 | comment | added | Richard Hardy | Thank you for the explanation. I just meant I do not get the grammar of that sentence. | |
| Apr 13, 2021 at 15:31 | comment | added | Demetri Pananos | @RichardHardy We're largely into semantics now, but if all there is to logistic regression is optimization of the log likelihood then this makes every deep neural net for binary classification a logistic regression. Maybe you're interested in perceiving it that way, but I am not. Linearity in the parameters is a part of logistic regression to me, so just because you optimize the bernouli LL does not mean you're doing logistic regression. But I repeat, we're into semantics about what is and is not logistic regression, and I don't care much for semantics. | |
| Apr 13, 2021 at 15:02 | comment | added | Richard Hardy | I do not get the first sentence of the comment above. | |
| Apr 13, 2021 at 12:15 | comment | added | Demetri Pananos | @littleO optimization of the bernoulli log likelihood does not a logistic regression make. OP's question was about statistical properties of these models and their ability to be tested via NHST. While modestly non-linear $f$ can be tested (e.g. splines), more generally deep neural networks make the problem of determining the null difficult. | |
| Apr 13, 2021 at 6:40 | comment | added | littleO | "Deep neural networks do not map nicely onto a statistical counterpart" In logistic regression we view the label $Y_i$ corresponding to the feature vector $x_i$ as a Bernoulli random variable, and we make the assumption that $P(Y_i = 1) = \sigma(f(x_i))$ where $\sigma$ is the sigmoid function and $f$ is a linear function. This is an arbitrary choice for the form of $f$, though. Couldn't we equally well assume that $f$ is a neural network? Then we're doing deep learning, and the optimization problem we solve is to maximize the log likelihood, just as in logistic regression. | |
| Apr 12, 2021 at 23:38 | history | edited | dimitriy | CC BY-SA 4.0 |
deleted 3 characters in body
|
| Apr 12, 2021 at 14:40 | vote | accept | Ryan | ||
| Apr 12, 2021 at 2:42 | comment | added | Demetri Pananos | I mean, there are many examples where it might seem practical; not all ML problems are solved with billion parameter models. The problem for testing is that a) its hard to derive what the null should be in a lot of cases, and b) it doesn't really matter because inference is not on the menu. Take penalized models for example. Its hard to know what the null would be for a penalized model, especially if it is for something like an SVM. Its hard, it isn't what we want, so we don't bother with it. | |
| Apr 12, 2021 at 2:21 | comment | added | Ryan | First, thank you for the well-considered response. Second, am I right in interpreting your answer, as well as one provided by a commenter above, by thinking that the sheer volume of data and number of features makes it impracticable to perform tests of statistical significance? Hence, we rely on a strategy of optimization for large data sets? | |
| Apr 12, 2021 at 0:59 | history | answered | Demetri Pananos | CC BY-SA 4.0 |