Nicole Mirea @nimirea@scholar.social

@davidbenque caveat: i'm usually talking to people who usually know what a probability distribution is

also i can explain backprop by just saying "to train this, we compare the model's output to the 'correct' answer that we have from the data, and adjust the connection strength between nodes in a way that will make that 'correct' answer more likely" because i'm talking to linguists and not computer scientists... they can ask me later about my optimization parameters 😎

@davidbenque oh! and before showing ANY of this, i start by explaining the higher-level purpose/task of the network: "a neural network is just a way to learn a function—that is, a mapping from inputs to outputs. in this case, the input is a phonetic symbol, and the output is a probability distribution over the upcoming phonetic symbol, or what the network thinks is likely to come next." i always start by talking about the input and output layers, then move on to the inner structure

@davidbenque so this is what I usually include in my talks. it's particular to my study, but the basic form will be similar. orange layers here are the recurrent layers, and the purple is an embedding layer.

i didn't draw all between-layer connections because that would make it hard to see, so i start by explaining that, & that each node is connected to every node in the next layer. i do a few separate slides to walk through it, with a box around the part i'm currently talking about

on a bigger note, I honestly think a lot of "deep learning"/"neural network" hype is due to its apparent inaccessibility. and that's more due to many people not explaining it well (understating limitations, touting it as a panacea), than it being inherently more complicated/good/smart

once you know that every machine learning technique has some limitation, you get better at finding it

most humans are really good at thinking in spatial terms!! give them the opportunity to succeed by using spatial metaphors! ("flow of information", "loop back", etc.)

today I tried to explain neural nets to people without visual aids

note to self: never try this again

@mpjgregoire it's kind of like, if you were a biologist, and you observed that some members of a species have different sets of mating rituals. you probably wouldn't call one mating ritual "wrong" (though it might certainly be maladaptive in some context). but you'd probably work to find out why the variation occurs, and what this variation depends on

@mpjgregoire as for "hav[ing] to be neutral about variants": it depends on the goals and the object of study. if the object is the human capacity for language, then each variant is a legitimate data point and MUST be treated as such.

@mpjgregoire Well, under those assumptions, we should all speak the same way in all contexts, in order to facilitate communication. But there's variation in language use, and that variation can be meaningful. This variation lets us do many other things with language besides state propositions: express an aspect of your identity, create closeness between yourself and another person, and many other things (some, implicit!)

@mpjgregoire maybe. we can't know, unless there's some documentation of his preference (AFAIK, there isn't).

i'm not interested in marking one way of speaking as "wrong", or in policing the way that people speak. i'm more interested in learning about and characterizing the variation in how people actually do speak..

I have so many results files. SO MANY. and I have no idea what most of them are, except that I know that the most recent ones were used in my write-up.

so what the hell do I do with the rest???

I'm going through an old (~1 year) codebase that I just haven't had time for, and, I can't stress this enough:

DESIGN YOUR CODE WITH A NAIVE DEVELOPER IN MIND. Yes, even if you're in academia; yes, even if you plan to be the only one working on the project.

Often, this will mean coming up with a plan for the directory structure *before* you even write a single line of code. This also means keeping track of the dependencies in a readme.

@glitternoodle ah! I'm more interested in the consonant at the end: is it "zh" (like at the beginning of "Jacques", or the end of "beige") or "z"

@sorreltyree oh! I'm more interested in the consonant, is it more of a "zh" sound (like at the end of "beige" or the beginning of "jacques") or a "z" sound

@mpjgregoire :) i don't believe there is a "correct" answer, but there's probably a "most popular" one, as well as factors (either implicit or explicit) that influence your choice of one or the other

also pls don't get nitpicky w/ me about maximum onset principle because it is honestly too early in the morning for the syllable discourse