From: Iain Strachan <igd.strachan@gmail.com>

Date: Mon Feb 12 2007 - 13:15:32 EST

Date: Mon Feb 12 2007 - 13:15:32 EST

On 2/12/07, Freeman, Louise Margaret <lfreeman@mbc.edu> wrote:

*> I don't see how stating a date in a dissertation can *not* imply endorsement
*

*> of that date! Is the same true of his central conclusion? Can you actually
*

*> write a 197-page scientific thesis and *not* believe what you've written to
*

*> be true?
*

*>
*

I beg to differ.

Actually, though I don't support the YEC ideas in any shape or form, I

DO, think that actually we do scientific research all the time based

on assumptions that we don't believe are true. Why is this the case?

Because the simplifying assumptions we make allow a tractable

solution.

For example, how many PhD theses (and scientific papers) are there out

there that have a linear regression trend plotted by "least squares

fitting"? How many people even know, let alone believe the underlying

assumption behind fitting a trend line (something we do almost without

thinking because Excel has a button to do it for you)? Here's the

assumption: you assume in making the fit that the data are best

modelled by the equation y = mx + c + r, where r is a residual that is

due to random processes (and errors in measurement). In choosing m

and c to minimise the sum of squares of the r variables, you are

assuming that the probability distribution or r is a Gaussian (aka

Normal distribution) whose standard deviation is independent of x.

Under this assumption, the least squares fit maximises the likelihood

of the data, and m and c are the "best fit". If the noise isn't

Gaussian, then m and c aren't the best fit and any predictions will

have a systematic bias.

Now, I'm guessing that many people who use linear regression don't

know this esoteric fact, so can't be accused of not believing it. But

to take another example: Hidden Markov Models are the subject of

many, many PhD theses out there (including mine). They are used

extensively in automated speech recognition. The model captures the

data as a sequence of discrete "states", in speech recognition each

state being a "phoneme" or component of speech. The underlying

assumption of the Hidden Markov Model is that the length of time spent

in each state (the time to pronounce each phoneme) is given by an

exponential distribution. No-one who uses Hidden Markov Models

actually BELIEVES this to be true. The true distribution is more

likely to be Gaussian. So why do we use a model that is based on

assumptions that we KNOW are false? Because (a) it works in practice

and (b) we can do the maths. If we had to model the states as

Gaussians, the resultant model would take many orders of magnitude

longer to process the data, and we wouldn't be able to do speech

recognition.

So actually I see nothing wrong or dishonest about doing science based

on assumptions that we don't believe are true - we always state "given

the following assumptions:", and we are not beholden in scientific

papers to state whether we personally believe those assumptions to be

true.

Iain

To unsubscribe, send a message to majordomo@calvin.edu with

"unsubscribe asa" (no quotes) as the body of the message.

Received on Mon Feb 12 13:15:42 2007

*
This archive was generated by hypermail 2.1.8
: Mon Feb 12 2007 - 13:15:42 EST
*