(aside image)

Matti Pöllä

Post-graduate Researcher, M.Sc. (Tech.)

Contact Information

Visiting address: Room B332
Department of Information and Computer Science
Konemiehentie 2, Otaniemi campus, Espoo, Finland.
Mail address: Aalto University School of Science and Technology
Department of Information and Computer Science
PO Box 15400
FI-00076 Aalto
Email: matti.polla@tkk.fi
Tel: +358-9-470 25115
Fax: +358-9-470 23277


See my publications page (available also as [BibTeX] [ps] [pdf]). See also Google Scholar, DBLP.

Research Interests

Artificial immune systems in data mining

Artificial Immune Systems (AIS) are computational models developed with the vertebrate immune system as a motivation. My latest research involves applying AIS models for data mining of natural language data and developing new AIS algorithms for text mining.

Modeling Anticipatory Behavior with Self-Organizing Neural Networks

Cognitive tasks performed by humans are often driven by anticipations about the future. Traditionally, the AI of an agent is implemented as a reactive if--then rule set, which allows the agent to behave only reactively. An internal predictive model can assist an agent to simulate future events and thus act anticipatorily. In my research the focus has been on building a prototype-based neural network estimate of the dynamic state space of the agent.

Sequential Learning and Forgetting in the Self Organizing Map

The Self-Organizing Map (SOM) is a visualization and clustering tool for creating a topologically correct mapping of a high-dimensional data set into a two-dimensional neuron lattice. The SOM is typically trained with a static set of data and thus all inputs are equally represented in the SOM projection. However, when training the SOM sequentially, representations of old inputs may overwritten, which can be understood as a manifestation of catastrophic forgetting found in feed-forward networks. Self-refreshing mechanism based on generating pseudo-data using the SOM codebook vectors can make the forgetting process a gradual instead of catastrophic.

Language Learning in Multi-Agent Systems

A group of autonomous agents operating in the same environment can benefit from inter-agent communication. The emergence of an shared language is based on finding a common semantic association between percept objects and utterances describing their properties. In the SOMAgent framework the semantic memory association task is implemented using the SOM.

Self-Organizing Map -demos

I have written some Java applets to demonstrate the Self-Organizing Map algorithm.

[1] [2]

SOM animation



I also have a personal home page.