User talk:Amir Mehrpanah

Score (statistics) machine learning section
"It might seem confusing that the word score has been used for, because it is not a likelihood function, neither it has a derivative with respect to the parameters. For more information about this definition, see the referenced paper."

I find this section confusing, specifically "neither it has a derivative with respect to the parameters", and don't think this information necessarily belongs in the score matching section. A clarification on the origins of the term score may be helpful, but I don't think such background belongs in the Applications section, nor does it help to explain why score functions are useful in ML. tiral (talk) 16:07, 12 July 2023 (UTC)


 * Thank you, @Zomboni13, for your feedback.
 * 1. Score matching is a technique in machine learning, which is used for generative modeling, please see the referenced paper if you need more information.
 * 2. This sentence, "because it is not a likelihood function", refers to the fact that likelihood functions, most of the time, have this form p(x|θ), which can be read as p(observations | hypotheses). Here, in score matching, we have p(x) or p(observation) which does not explicitly show that form to be called a likelihood function.
 * 3. In this sentence, "neither it has a derivative with respect to the parameters", I have tried to emphasize that the derivative is taken wrt x and not parameters θ. It is important to note that the derivative is taken wrt to a hypothetical location parameter. This also can be found in the referenced paper.
 * 4. Do you have any suggestions about where to put score matching if not under applications?
 * 5. Unfortunately, I'm not really sure why it is called score function, while it takes some time to absorb how it can be seen as a score function. Amir Mehrpanah (talk) 16:26, 12 July 2023 (UTC)
 * 1. I am aware that score matching is a technique for ML, I am making the point that the section, as written, made no direct reference to this fact, nor did it even define score matching. I have since added a short paragraph that can be expanded on.
 * 2 + 3. I am fine with this, my confusion is more so in the language of the paragraph than in the content. Note that, because this is a statistics article, none of these functions are defined to have parameters (learned or otherwise), so if you assume them to have parameters like coefficients to be changed then I think that should be notated. I did this in my edit with s_\theta.
 * 4. Note that I say "A clarification on the origins of the term score may be helpful, but I don't think such background belongs in the Applications section". I am not proposing that score matching be moved, but am proposing the context on why score functions are called such be moved out of Applications. Score matching should stay where it is.
 * 5. I thought this was the point of your reference? tiral (talk) 17:00, 12 July 2023 (UTC)
 * Sorry, I myself have bad language in my point on 2+3. My point is that the sentence is confusing if you say that you're taking a spacial gradient \nalba_x \log p(x) when no term in the equation has a parameter variable explicitly expressed. tiral (talk) 17:02, 12 July 2023 (UTC)
 * 1. Thank you for making it clear and giving a bit more context.
 * 2+3. You are right, it might be better to expand a bit more on this fact, and clarify the hidden location parameter in the definition of the "score" as done in score matching.
 * 4. I don't have any specific page in mind dedicated to score matching for generative modeling to explain the origins of this word and how it connects to the definition of score in statistics.
 * 5. I mean, it is indeed possible to see it same as the score defined in statistics, but it needs a bit more time to absorb and understand. I myself was searching for a few days, and finally I found the solution in a paragraph of the referenced paper. So I think it might be helpful for others that ask "why is it called such?" then, they, like me, search score on Wikipedia and find some useful references. But I guess it will be fine if we manage to address points 2-4.
 * Thank you for your time @Zomboni13. Amir Mehrpanah (talk) 17:39, 12 July 2023 (UTC)