precalculus does matter

[trochee]

I just spent three days working out why two different language modeling schemes were producing vastly different log-probabilities. I was using every bit of good software design practice that I know, and leaning on colleagues to confirm that I was running it right.

You know why it all went wrong? System X produces probabilities in logarithms of base 10, and system Y does it in base e.

This reminds me of the stories about the Mars Climate Orbiter unit conversion error, although that is apparently apocryphal.

Comments

[evan.livejournal.com]

shouldn't log_e and log_10 just be off by a constant multiplier?

[trochee.livejournal.com]

yup, they're different by a constant multiplier of log[e](10). But when you're log-linearly interpolating between the log values it matters.