The question I wanted to answer for myself is why we clearly can read words "all at once" at a glance, and yet the order of the letters in the word matters. That seems perhaps like a silly question ... but what would be an algorithm that would run in parallel on parts of the word, completely independently, and yet somehow deal with the implied word ordering. I hypothesized that by grouping the letters of the words into pairs there would be a way to derive the original word - even if it were misspelled.
And it works well, even though there is no inherent reason it should. One would expect it would require a deep neural network, but this is insanely shallow.
In my big dictionary of 370,000 or so words, only 255 don't return themselves as the correct word. In my 49000 or so word dictionary only 10 or 12 don't pass that test.
That's amazing!
It supports my hypothesis that people process words in letter pairs, in parallel, at a glance.
No comments:
Post a Comment