Home

Previous Entry | Next Entry

note to self

  • Feb. 27th, 2008 at 12:43 PM
cat - after teo rus, bad post!, sleepy goth, crossroads, money, mikage: go deeper, treason and plot, raahhhr!, O RLY? YS RLY!, asuka, bitch please, Angry Young Meredith, fat bob's novel, rose window, space hog!, sup fruits, oh really?, me!, gunpowder, doom, needs head!, native habitat, dehydrated water, brave, serious
When reading papers about math to translate into code, avoid flipping the fuck out about how difficult X will be to implement until you have verified whether the paper's approach actually uses X.

"However, in place of the expensive Fisher kernel1, we present a new sequence-similarity kernel, the spectrum kernel..."

1Which happens to be chock full of partial derivatives and other things that are no fun to implement in things that aren't Matlab.

Phew.
  • Mood:sheepish

Comments

( Leave a comment )
[info]jered wrote:
Feb. 28th, 2008 03:36 am (UTC)
Wow. I have absolutely no idea what you're talking about. This is a relatively novel experience.
| Reply | Thread
[info]maradydd wrote:
Mar. 1st, 2008 05:08 pm (UTC)
More support-vector-machine foo. A kernel function K(x,y) is used in classifying an unlabeled point; once you've trained your classifier and have your list of support vectors, classifying a new point is simply a matter of summing ciK(x, xi), where xi is a support vector and ci is the corresponding class label (typically +1 or -1).

The Fisher kernel is pretty gory, at least if you're me and could really use some remedial tutoring in both partial differential equations and matrix algebra.
| Reply | Parent | Thread
( Leave a comment )