July 21, 2019
MarkBernstein.org
 
Follow me on Twitter

Neural updates

I’ve been exploring some neural networks for helping Tinderbox automatically classify some notes.

For example, some Tinderbox notes inevitably describe plans you’re making:

I thought a neural net might be able to sort these planning notes out from all your other notes, and that might be handy. So far, we’re right about 7 times in 8, which isn’t too shabby.

Yesterday, I tried extending this to also recognize notes that review a book, article, play or movie — that offer an opinion or recommendation, or that summarize the work. About ⅓ of the items in this weblog are, in this sense, reviews. I trained the existing net with about 50 extracts from the weblog, and — what do you know — it’s astonishingly good. Plenty of mistakes, but it does the job. Oddly, the classifier seems to have an easier time recognizing reviews than recognizing plans.