Ian

For a while I’ve wanted to train a neural net on my handwriting, as part of a larger test project that involves a drawing interface. To minimize error effectively I need the highest recognition rate possible, so using a generic dataset like IAM for testing is not necessarily the best. Plus, I don’t know enough about machine learning currently to entertain adaptive networks, that start from some generic test set and gradually adapt to the writer’s writing style. 

I used to contemplate developing an HTML5 app connecting to a backend that records writing on my iPhone, but although that allowed for on-line input, it was less straightforward for a hobby project. An alternative was to develop a template paper with a grid and scan them, but this was also time consuming. Luckily, I have a ReMarkable tablet to speed up the process!

The ReMarkable has a grid template that makes it easy to enter characters en masse:

The ReMarkable desktop app lets you easily export this writing to a folder of PNGs. Within two days I already have, for instance, 329 images of the letter “x” in my handwriting (extracted as separate images of characters using a Python script). My task does not necessarily require word input (yet), so I’m going to make do with what I have and build the model from here.

From Networking Peripheries by Anita Chan. The emphasized statement is actually incorrect and puts forward a common misconception perpetuated by “technorealist” circles: in fact, constructionism is not the kind of constructivism Papert rebelled against, where students should self-educate. Here, in writing Chan betrays a technocentric assumption that equates Negroponte’s ideas to those who work(ed) in the lab he founded, flattening the “rural” as a privileged untapped source of disruptive or inconvenient knowledge and the “urban” as a spigot from which techno-solutionism flows, forgetting that disruption is not just (always) on the periphery: it can be front-and-center, operating not merely through what is spoken but what is *heard*, how things are taken, translated, and misconstrued across conglomerations of multi-national actors with varying aspirations and political motivations. The silence (or un-examination) of those who may contend with OLPC within its very womb is not actually proof of techno-solutionist homogenization. 

Lately I have been learning the (extremely lengthy and disciplining) process of turning “jottings” into full field-notes. After spending over a day converting only one four-hour session, I’ve realized that much of the process’ tedium is marred by manual expansion of shorthand.

Shorthand has been around for centuries, consisting of (sometimes incomprehensible) ticks and slashes. Typing on computer makes it slightly different: I’ve adopted my own shorthand based on disemvoweling, IM abbreviations, and making abbreviations for local common words like “Swahili” (it’s Sw). I’ve also adopted a convention whereby gestures are placed in-line with words using (when I’m good) double-brackets e.g. [[points to board]]; and a naming convention where commonly-used names are abbreviated as capital first letters.

Facing the daunting task of converting more jottings, I wrote a simple Python program that reads a text file, finds the words and then converts any shorthand it finds into longhand with a dictionary lookup. This alone has already saved me tons of time, for those shorthand abbreviations I can be sure won’t be confused with other things. 

However, one can go further... For instance, the abbrev. “ur” for “you’re” and “your” makes it impossible to use simple conversion; a full NLP parser would infer which is most accurate. (This is probably an easy task for any expert NLP practitioner). In another instance, I spend time isolating the gestures and names in dialogue to full, properly-written lines, like ‘“Hello X,” said M, pointing to X.’ The ethnographer can then work with these expanded jottings when making field-notes, reducing a lot of tedious effort in the process.