Fullscreen
Loading...
 

Print

ISC-PIF Guest Researcher  //



Taras Kowaliw



Image
B.Sc. (Mathematics; Toronto), M.Sc., Ph.D. (Computer Science; Concordia),

Guest Researcher / Chercheur Post-Doctorant

Institut des Systèmes Complexes - Paris île-de-France,
Centre national de la recherche scientifique,
57-59 rue Lhomond, 75005, Paris, France
t.: +33 01 42 17 40 35
f.: + 33 01 45 35 79 21
w.: http://kowaliw.ca
e.: firstname . lastname @ iscpif.fr








Research


Image
My work involves automated design: that is, the creation of specifications for structures, networks, or programs via machine learning. Due to its capacity for finding solutions in situations where the appropriate structure or complexity is unknown, I concentrate chiefly on evolutionary design. My colleagues and I have developed means for generating programs capable of performing computer vision tasks, generating structural forms, and generating electronic art suited for individual viewers. The results are often highly robust: our work in computer vision is capable of extracting good features directly from several diverse image databases, including those designed for object recognition, artistic style recognition, and medical cell classification.

> Three automatically discovered transforms highlighting important features in an image (top left).


Image

A key issue with evolutionary design, as with other techniques, is that search often becomes bogged down with increased pattern size. It has been argued that in order to achieve the sorts of complexity, robustness, and adaptability we would like, engineering needs to adapt the methodologies of complex systems science. Of particular interest to me is the use of techniques inspired by embryogenesis. In nature, embryogenesis increases the evolvability of forms known useful in solving these problems as faced by multicellular organisms.
 


< Evolved artworks. Individuals consisted of a multi-agent growth in a pixel-based world. They were selected by users based purely on their aesthetic preferences.


The new field of Artificial Development (AD, sometimes Artificial Embryogeny, or Computational Development, etc.) deals with the use of metaphors from biological embryogenesis as a design step in an automated search. It is hoped that a developmental stage between representation and pattern will allow for the incorporation of these same capacities for increased evolvability. At times, this is due to a good fit between the developmental model and the associated problem. For instance, the use of multi-agent embodied systems tends to generate patterns which are repetitive, fractal, and highly discontinuous, traits useful for generating electronic art. Another aspect of AD is that it leads to a dissociation between the sizes of the representation and pattern, sometimes providing the ability to evolve larger patterns via a simpler genotypic space. Development also allows for the incorporation of environmental information in the design stage, meaning that the information in a particular pattern comes from both genetic and reactive components. This allows for a sort of artificial polymorphism, where a single genome can lead to several different phenotypes depending on the particularities of the developmental environment.
 

Image
Several truss designs, each grown from the same genome, but in one of several different environments.

 

My chief interest at present is in finding means of incorporating and analyzing various strategies for development into automated design techniques. I am currently developing a open software package to implement this.

 

Selected Publications

 




<< back to Researcher's page

Contributors to this page: taras and webmaster .
Page last modified on Wednesday 04 May, 2011 18:23:42 by taras.