TE
TechEcho
Home24h TopNewestBestAskShowJobs
GitHubTwitter
Home

TechEcho

A tech news platform built with Next.js, providing global tech news and discussions.

GitHubTwitter

Home

HomeNewestBestAskShowJobs

Resources

HackerNews APIOriginal HackerNewsNext.js

© 2025 TechEcho. All rights reserved.

Reading Chess (1990)

38 pointsby slbenficaalmost 7 years ago

2 comments

cmontellaalmost 7 years ago
I took a couple courses by Baird at Lehigh before he retired. He used to teach &quot;Pattern Recognition&quot; as it was called before &quot;Machine Learning&quot;. His approach was decidedly old school, but really that was the whole point; a lot of the trends in computing are cyclic, and if you focus on the fundamentals you have context for where new trends fit in.<p>One of his biggest take-aways in building systems like the one in the linked paper was the need for layered understanding. For example, to understand the meaning of a passage of scanned text, you can&#x27;t just look at each individual letter of the text. There is ambiguity in characters like l and I or 0 and o and O. To remove that ambiguity, you bring in more context by looking at a character in a word. Then a word within a sentence. Then a sentence within a passage. This is how they were able to build highly accurate systems without massive amounts of training data.<p>Now that I&#x27;m on a roll recollecting the teachings of Henry Baird I will add two more:<p>1) There is a difference between an algorithm and a heuristic that is often overlooked these days: algorithms are provably correct. If you can&#x27;t prove it&#x27;s correct, then it&#x27;s a heuristic. By this measure, there are very few algorithms out there, because there are very few proofs of correctness. Most likely, anything you&#x27;ve ever called an algorithm that you&#x27;ve written yourself is actually a heuristic.<p>2) Because of the above, Baird was very fond of KNN classifiers, because they are one of the few classifiers that have any sort of provable guarantees -- namely that with infinite data, the error rate is guaranteed to be no worse than 2x the Bayes error.
Nzenalmost 7 years ago
tl;dr web rendered case study of a computer vision system by K Thompson and H Baird. They used ocr plus word level heuristics to digitize an encyclopedia of chess games&#x2F;moves. By modeling each game during a second page of parsing, they were able to constrain the possible interpretations of questionable characters&#x2F;words to legal moves. They claim this reduced their error rate 30 fold.