Allison Parrish is a user on You can follow them or interact with them if you have an account anywhere in the fediverse. If you don't, you can sign up here.

Allison Parrish

I logged into the everyword twitter account a few weeks back to set up 2fa and didn't realize that the act of adding a number to the account would make twitter believe it had license to start texting me all the time. weirdly the only account it would just send me notifications about was... everywordcup?

misread this as "most accurate source of hypothetical weather"

excerpt from the beginning of L.M. Montgomery's Anne of the Island after being resized to exactly 50000 words (using scipy.ndimage.zoom on an array of word vectors)

amazon's aws product line is remarkable in its needless breadth and in how so many services have overlapping functionality nearly to the point of being duplicates

experimenting upscaling text with convolutional neural networks, latest update: I can now pass in random character ngrams and get back seemingly unrelated text that looks... vaguely... estonian?

visiting someone's campsite who decided to put tom nook behind bars where he belongs

always somewhat surprising to be reminded that ecclesiastes is part of the christian canon

more text super-resolution experiments with neural networks, this time using bigrams instead of words as the underlying unit. (this is after interrupting the training like 5% into the first epoch just to see what it looks like)

tbh kinda jealous of the research that these other allison parrishes are doing

putting this neural network through its paces by trying to predict which (actual) words sound most similar to strings of random characters

t-sne of the character embedding layer weights from my orthography-to-phonetic-similarity-vector neural network. reading t-snes is like reading tea leaves but there are some interesting and encouraging clusters here (e/a/i, c/k with back vowels o/u, m/w)

training a neural network to predict phonetic similarity vectors for words based solely on their orthography (i.e., not in cmudict). the neural network is only at about 20% accuracy but the results are already promising! showing the most similar word in cmudict here just as a sanity check... (context:

i can’t stand the single storey a in the ios 11 notes app. what a bizarre and annoying typography design decision. it’s distracting to the point of making the app unusable

though there is definitely something distinctly cubist about that image and the whole idea of breaking faces down into components, reminds me of