locally minimized.

I lost the blog post that I was reading by Some Guy about Some Thing, but I would like to digest it a bit here regardless. The Guy I will definitely not be able to remember, but The Thing I can flesh out still. I think. The good thing about losing source material is that I don't have to worry about getting facts right: I can worry more about what feels true than what is real. 

But digression! 

The article I was reading was about machine learning and backpropagation, which is (and don't murder me on technicalities here) a way to tweak a neural net's weights based on a dataset to best solve for new inputs. Basically think of a bunch of strings attaching a network of bells, and we're trying to make it that when we yank on a starting string the bells produce a meaningful response. Backpropagation is the act of "strengthening" sections of the string so that they impact the final result in a way that better matches a set of data. This is probably a terrible analogy, but it's the best I got right now. The MORE IMPORTANT thing about this how we go about adjusting the strings in the first place, which is usually done with a gradient descent algorithm. This is just a fancy way to say that we adjust things in set pattern of steps to have the whole bell apparatus settle into a solution that works for what we have observed. 

But this solution can be a solution that isn't optimized to the most broad view. Depending on the steps taken, we can settle into a localized solution versus something more global. Imagine if I told you to find the lowest point in a city, but required you to close your eyes and then run 20km between your observations. Each step in this case, 20km, would set you in a new place of the city where you could observe which way is downhill and then run 20km in that direction (eyes closed, remember). In this fashion you probably will run over the lowest point again and again, never reaching a minimum of any kind. On the other side of things, if you were to take a step size of a foot, you'd end up in a slight sinkhole somewhere in the city. Standing in the sinkhole, walking a foot out of it, noticing downhill is back into the sinkhole and concluding the sinkhole is the lowest point in the city. 

There's ways data scientists get around this by adjusting step sizes as they go along. BUT my train of thought is more around this idea of finding solutions to my surroundings in a more general sense. Solutions can be things like "Where can I find a good Taco?" or "What's the best place to take someone named Gillian on a romantic date?". To find a solution to this involves my own experience and then my ability to find and integrate new information. And how I receive information, the mechanism that transfers this information, is a type of step size. If there's a good friend of mine that only eats a certain type of food, that aren't tacos, she probably won't be a good person to ask about tacos. The step sizes she would afford me to find good tacos would have me bouncing all over a taco landscape with little hope for finding global values. BUT, maybe she's a romantic at heart and the step size she would give me for finding a place to woo Gillian, would be the perfect amount of petite. 

People are a bit harder to think about as having step sizes, because people change the way they present information depending on topic. An easier step size disseminator to think about is social media.

I often think about the reasons I never have liked being on social media. I think one of the main reasons is I have a hard time representing myself in a way that feels authentic when on a platform that requires specific types of content to be created. Square photos. Short bursts of text. Whatever. I've realized also that social media creates a very small step size for me that makes me feel that I have gotten stuck repeatedly in a local minimum (or maximum) while being led to believe I'm in a global minimum (or maximum). I guess this is a reframing of the echo chamber everyone always talks about, but I don't think that that captures very well what I'm speaking of since I know of people that can make social media perform with amazing step sizes for them. These people I think of a bit like someone that can put a perfect bullnose on a piece of wood with a chisel: I wouldn't think it possible, but god damnit they can do it.

In some ways these people are finding the diagonals when the rest of us are just bouncing up, down, left, right on a grid. They're cutting corners and finding shorter paths and better solutions (https://en.wikipedia.org/wiki/Power_series). Good tools ("tool" being used in this case as some apparatus that distills and organizes an environment... which I think social media can be) should always be finding the square root of 2 instead of just 2. But maybe that also requires WHERE you want a tool to take you. 

I guess the big point that I've been thinking: we get information through tools, and these tools have step sizes: chunks they break the world into (Conversation is a tool! And conversation in non-adjustable step sizes is rhetoric: a tool used to achieve a sense of local stability without hope of continued growth.) These chunks can have default sizes that don't suit us: they lead us to only local solutions when we could be reaching for something more encapsulating. If we are thoughtful we can put a tool into a different context to better find local maximums and minimums. But it requires intent.

I have a story of small step sizes that involves high heel shoes. I was at a dinner party and a woman announced that she had always wanted to walk on someone while wearing high heel shoes. I offered myself up as a test subject and can report that being walked on by someone wearing high heel shoes is extremely uncomfortable. However, in such small step sizes I do believe I was shown a global maximum of someone's thought process. So as a step sizes go, those couple inches at a time seemed like miles.