Machine learning
Nov. 18th, 2011 07:39 am![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
lederhosenThing 1: I've been doing the free machine learning e-course from Stanford. I can thoroughly recommend it* and if you missed it there's another offering in January.
Thing 2: Stanford are also offering a whole bunch of other free online courses. The PGM one looks like it might be work-relevant, and I think I might do the game theory one on my own time.
Thing 3: The course forums are pretty good, in general, but there's this one guy who makes me cry every time I see his comments. In this case, the OP produced a graph that looked good near the fitting points, but produced negative results (nonsensical in context) elsewhere**. Response, in the original punctuation:
Translated into English: "You should censor the results that are obviously wrong, to make the graph look better, without considering what those results might imply for the accuracy of the other results that don't LOOK wrong."
...and that, right there, is a "I would never ever hire you" button.
*although I still don't like their explanation of back propagation - not sure whether I've misunderstood it, or whether it's just a bad explanation of something that should just be "calculate the gradient of the cost function".
**Kids, JUST SAY NO to extrapolating from eighth-order polynomial fits. Please.
Thing 2: Stanford are also offering a whole bunch of other free online courses. The PGM one looks like it might be work-relevant, and I think I might do the game theory one on my own time.
Thing 3: The course forums are pretty good, in general, but there's this one guy who makes me cry every time I see his comments. In this case, the OP produced a graph that looked good near the fitting points, but produced negative results (nonsensical in context) elsewhere**. Response, in the original punctuation:
HOLY FLOW CHARTS....I have never have seen negative Water Flow Across a DAM... When the Level Transmitter shows a ZERO LEVEL....you only get SPLASH flow across the Dam (wind and waves cause Splashing). A secondary flow sensor should pick up this random flow. When the Level Signal goes positive you should get a flow signal that follows Weir Flow Charts...I think it is a 3/2 or 5/2 power or something in between depending on how it is constructed.
Modelling it with a Chart showing NEGATIVE FLOW would get you FIRED...recommend erasing everything < than "ZERO" FLOW...the Stanford computers might pick that up as an ERROR in the CALCULATIONS...
Translated into English: "You should censor the results that are obviously wrong, to make the graph look better, without considering what those results might imply for the accuracy of the other results that don't LOOK wrong."
...and that, right there, is a "I would never ever hire you" button.
*although I still don't like their explanation of back propagation - not sure whether I've misunderstood it, or whether it's just a bad explanation of something that should just be "calculate the gradient of the cost function".
**Kids, JUST SAY NO to extrapolating from eighth-order polynomial fits. Please.
