Re: Backpropagation Neural network Code problem

On Mar 27, 9:15 pm, Greg Heath <he...@xxxxxxxxxxxxxxxx> wrote:
12345678901234567890123456789012345678901234567890123456789012345
On Mar 26, 11:49 pm, "Adeel " <neoresearc...@xxxxxxxxx> wrote:

Could I said it to you please sir download thedata
yourself from the net &#8220;BioIDFace> Database&#8221;
featured extracteddata.

Preferable to email the data in ascii format and reference
the URL for contextual info.

The &#8220;BioIDFaceDatabase&#8221; had 1520 total records.
60 % of it means 919 records. ( I am dividing the total records
in 60 % for training and 40 % for testing). (And yes I did care
how to divide all the records in 60 40 i.e. some person pictures
are more some had less so I divide accordingly to the individual
person and not ot the whole dataset.)

How many for each person(class)?

Oh boy!

I was finally able to get these training set
MSE results using all of the data and a
I-H-O = 40-41-5 MLP with

Nw = (40+1)*41+(41+1)*5 = 1891
Neq = 1520*5 = 7600
r = Neq/Nw = 4.02

Nepochs MSEtrn
1000 0.072
3000 0.00676
4300 0.00229
8300 0.000796

As a reference,

MSE00 (Assumes the output is equal to the mean
of the training set) = 0.229

MSE0 (Backslash Linear classifier) = 0.1073

However, this is supposed to be a classifier.
Therefore, the only practical performance
descriptor is percent classification error
or correct classification rates... Both 5
class error rates as well as an overall
mixture error rates.

In order to do this I took a look at the 5-dim
output matrix,t, expecting to see 1-of-c coding
for 5 classes (persons).

Instead, I find 23 different target IDs coded
via 5-bit binary.

Perpexed, I found the BioID Face Recognition
database and found that there really are 23
faces with their IDs coded in 5 bit binary!

I am sure that it will be better to convert
the 5-bit binary target matrix to one of 23
23-bit binary. Otherwise you will have to do
the inverse conversion implicitly within the
code.

Have mercy on your hidden units and change
the output coding.

Keeping H = 41

Nw = (40+1)*41+(41+1)*23 = 2647
Neq = 1520*23 = 34960
r = Neq/Nw = 13.2

P.S. What is the goal of your master's thesis?
.

Relevant Pages

  • Re: GCD(0,0)
    ... >quasi wrote: ... >>>Yes, it's fair, but it's also incorrect. ... >>>which divide into a and b, not numbers that a and b divide into. ... I challenged that, asking for a reference. ...
    (sci.math)
  • Re: binary divide by 3
    ... >> I vaguely recall seeing a reference to an efficient algorithm for ... >> performing a divide by 3 on processors that only have boolean ops, ...
    (comp.arch.embedded)
  • Re: complex orthogonal group
    ... > But if you mean that there is some kind of spectral theorem for ... I cannot remember how to reduce the classification for SO_n ... about so_n IS reduced to the pencil problem, ... reference with a proof of the lemma (in proof of which I used the ...
    (sci.math.research)
  • RE: Hello to Perl World (U)
    ... Classification: UNCLASSIFIED ... reference. ... Hope to be posting my queries in future. ... As for starters I would like ...
    (perl.beginners)
  • Re: goofy frequencies
    ... I'd like to have the option of phase locking it to some standard ... nominal 10 MHz result against the reference 10M and close that loop. ... The DDS chips are spendy, granted, but you didn't mention ... Divide by 61083979 and reference against 2Hz, ...
    (sci.electronics.design)
Loading