Unwise thresholds

[torkell]

I think I need to change the error threshold in the neural net I'm writing. After 7,548,600 iterations the error is 0.4069111311875221. I don't think it's going to reach 0.01 (the threshold I had set) in the near future (it's been running for over 16 hours). Time to change what the criteria for "problem solved" is.

In related news, Excel appears to only draw 32,000 points when told to create an X/Y scatter graph:


The initial error was 1.000167343112047, but I've trimmed the scale to try and show the curve.

Comments

[ralesk.livejournal.com]

:O

You’re listening to Fantasy Cargo?

[torkell]

*checks* Yep.

I'm impressed that you got it just from the "go".

[ralesk.livejournal.com]

That’s the only Akerheden song I have. I’m impressed that someone else has it! :)

And I wish I had had broadband before TiS went down :/

[torkell]

I have a few more Akerheden songs if you're interested - I think I got them from his site, which I probably found via the wayback machine.

[ralesk.livejournal.com]

Shame most of TiS is unaccessible through WM :(

[ralesk.livejournal.com]

You sure that the limes of that graph (iterations to infty) is nought? Who knows, it might be heading for $0.4$ or $0.{\dot 3}$, for all we know :P

// screw Unicode for not having a number-ready combining dot

[torkell]

You've got a point. It looks like it's going to bottom out at around 0.4069. Unfortuantly I don't know if that network was correct, as I killed the process and didn't write a way to dump the network without exiting.