## Re: [Artificial Intelligence Group] how long it takes to train

Expand Messages
• Regarding using a 5 bit number to represent 26 letters of the alphabet as neural network output representation: Kenneth Bull wrote: Why are you using 26
Message 1 of 5 , Mar 18, 2003
• 0 Attachment
Regarding using a 5 bit number to represent 26 letters of the alphabet
as neural network output representation:

Kenneth Bull wrote:
"Why are you using 26 units? Why not use 5 units to represent a binary
number between 0 and 31 instead of 26 for 0 to 25? (A=00000, B=00001,
C=00010, ..., Z=11010)"

This would require the neural network to learn a much more complicated
mapping for each individual bit- in essence learning the clasification
and learning to be a 32-to-5 encoder. Consider your bit number 4 in
the above examples you've given. Classes "A" and "B" have bit number
4 (counting from the left) the same (zero) and classes "C" and "Z"
have it as a one. This is arbitrary and probably does nto reflect
structural differences in the classes.

This sort of representation is suggested frequently online but is
dreadful in practice since the model needs to learn mappings to
classes and also learn to turn off those mappings in a convoluted way.

Keep in mind that the basis functions used by most artificial neural
networks in practice are simple, monotonic transfer functions.
Learning to turn "on" one of your bits requires one or more basis
functions. Learning to turn them "off" for this arbitrary
representation may take many more.
• hi all sorry replying late, bcuz of my exams. and thanks to predictorx and kenneth bull for replying. but my question remains the same and unanswered. is my
Message 2 of 5 , Mar 27, 2003
• 0 Attachment
hi all

sorry replying late, bcuz of my exams. and thanks to
predictorx and kenneth bull for replying. but my
question remains the same and unanswered. is my
network is effecient enough to learn alphabets and how
long it will take to train.

Touseeef
> Regarding using a 5 bit number to represent 26
> letters of the alphabet
> as neural network output representation:
>
> Kenneth Bull wrote:
> "Why are you using 26 units? Why not use 5 units to
> represent a binary
> number between 0 and 31 instead of 26 for 0 to 25?
> (A=00000, B=00001,
> C=00010, ..., Z=11010)"
>
>
> This would require the neural network to learn a
> much more complicated
> mapping for each individual bit- in essence learning
> the clasification
> and learning to be a 32-to-5 encoder. Consider your
> bit number 4 in
> the above examples you've given. Classes "A" and
> "B" have bit number
> 4 (counting from the left) the same (zero) and
> classes "C" and "Z"
> have it as a one. This is arbitrary and probably
> does nto reflect
> structural differences in the classes.
>
> This sort of representation is suggested frequently
> online but is
> dreadful in practice since the model needs to learn
> mappings to
> classes and also learn to turn off those mappings in
> a convoluted way.
>
> Keep in mind that the basis functions used by most
> artificial neural
> networks in practice are simple, monotonic transfer
> functions.
> Learning to turn "on" one of your bits requires one
> or more basis
> functions. Learning to turn them "off" for this
> arbitrary
> representation may take many more.
>
>
>
>
>
>
>

__________________________________________________
Do you Yahoo!?
Yahoo! Platinum - Watch CBS' NCAA March Madness, live on your desktop!
http://platinum.yahoo.com
• This will depend, of course, on many factors, but assuming that you ve got relatively current hardware, I d say something was wrong if most classes couldn t be
Message 3 of 5 , Mar 28, 2003
• 0 Attachment
This will depend, of course, on many factors, but assuming that you've
got relatively current hardware, I'd say something was wrong if most
classes couldn't be distinguished accurately in less than an hour.

One difficulty in this problem is that 26 separate outputs are being
trained simultaneously, which permits the possibility that some will
finish training before others. Consequently, modeling of some classes
may be overfit while others are underfit.

"Touseef Liaqat" <paramount01us@y...> wrote:
"i am doing a project on character recognition and i am using
back-prapogation algorithm for training the net. sample data set is
consist of bmp pics of characters. each character size is 5x7 pixels
so the input layer contains 35 units. i first trained two characters
so the output layer has 2 units and i have made only one hidden layer
which has 20 units. this network works well and trained with noise
data. but things become worst when i increse the output units with 26
for all alphabets and hidden layer to 30 units. this new network is
not trainning with all other parameters are same. i waited 2 to 3 days
for its training but nothing is happening.

my question is that how long it takes to train the network of this
kind(input units=35, hidden units = 30 , output units =26)?"
Your message has been successfully submitted and would be delivered to recipients shortly.