costapan.webpages.auth.gr Open in urlscan Pro
155.207.1.6  Public Scan

Form analysis
0 forms found in the DOM

Text Content

CONSTANTINOS PANAGIOTAKOPOULOS

I am a Professor at the School of Technology of the Aristotle University of
Thessaloniki.
 
Papers concerning my research in Theoretical High Energy Physics and Cosmology
can be found here .
 
My research in Machine Learning is the following.

--------------------------------------------------------------------------------

Margin Perceptron with Unlearning


Margin Perceptron with Unlearning

Ref: Panagiotakopoulos, C., Tsampouka, P.:The Margin Perceptron with Unlearning.
ICML (2010) 855-862

Solving for hard margin (L2-soft margin)

Download source code     v 1.1   Feb 2013   exe


Solving for L1-soft margin

Download source code     v 1.1   Feb 2013       Single precision source code    
v 1.1   July 2013   exe exe

Perceptron with Dynamic Margin

Ref: Panagiotakopoulos, C., Tsampouka, P.: The Perceptron with Dynamic Margin.
ALT (2011) 204-218


Download source code     v 1.1   Feb 2013   exe

Margitron

Ref: Panagiotakopoulos, C., Tsampouka, P.:The Margitron: A Generalized
Perceptron with Margin. IEEE Transactions on Neural Networks 22(3) (2011)
395-407


Download source code     v 1.1   Feb 2013   exe

Perceptron with Weight Shrinking

Ref: Panagiotakopoulos, C., Tsampouka, P.: The Role of Weight Shrinking in Large
Margin Perceptron Learning. arXiv:1205.4698 (2012)

Perceptron with Constant Shrinking

Download source code     v 1.0   Feb 2013   exe


Perceptron with Variable Shrinking (with n=3)

Download source code     v 1.0   Feb 2013   exe


Stochastic Gradient Descent

Ref: Panagiotakopoulos, C., Tsampouka, P.: The Stochastic Gradient Descent for
the Primal L1-SVM Optimization Revisited. arXiv:1304.6383 (2013) (accepted at
ECML/PKDD 2013)

SGD-r

Stochastic gradient descent with random selection of examples

Download source code     v 1.0   Apr 2013       Single precision source code    
v 1.0   July 2013   exe exe

 



SGD

Stochastic gradient descent with single (l=1) or mixed single-multiple (l >1)
updates and relative accuracy epsilon

Download source code     v 1.0   Apr 2013       Single precision source code    
v 1.0   July 2013   exe exe

Perceptron with Dynamic Margin
Margitron
Perceptron with Weight Shrinking
Stochastic Gradient Descent

Remark

In the above programs the seed of the random number generator was fixed to 0
which was the default value of previous Cygwin releases.

Instructions

The programs compile with the g++ compiler.
In order to make the .exe under Cygwin type the command:
$ g++ -Wall -lm -O3 file.cc -o train
 
To extend the maximum amount of allocatable memory set the desirable size in the
.exe file. E.g., for a size of 1024 MB the command is
$ peflags --cygwin-heap=1024 train.exe
For the .exe files given the heap size was set to 1024. To run the .exe files on
Windows platform one needs cygwin1.dll which comes with the Cygwin setup.
 
To see the available inputs for each program write
$ ./train
 
To run the program write
$ ./train [inputs] datafile modelfile
 
The datafile should be given in SVM-Light format. This means that each example
takes up one line. The label is from the set {-1,+1} and comes first. Then only
the attributes with non-zero value should be provided separated from their value
by the character ':'
A typical line reads like this:
-1 1:2.11 3:4.01 7:9.0 15:2.5
 
If the user doesn't provide a name for a modelfile one will be created in the
form datafile.model. The modelfile contains the components of the produced
augmented weight vector. Especially, for the perceptron with unlearning solving
for L1-soft margin the weight vector a is divided by b since w=a/b . Always the
first component corresponds to the extra feature rho of the augmented patterns.

Contact

For any question regarding the papers or the programs feel free to contact
either of the authors.
emails:   costapan@eng.auth.gr     petroula@auth.gr