Gábor Szirtes
http://people.inf.elte.hu/gszirtes
ELTE Faculty of Informatics
Dept of Information Systems
NIPG group

Motivations

Remote starting: what makes an information system ‘efficient’? (TD approach!)

A hint: let’s take a look at the Internet!

Continuously renewing vast pool of information (environment)
Several routes of acquiring infos
Locality
Noise resistence (structural and data(input) noise)
Variability

The matrix has you

The biggest information pool
Continuous change in space and time
Unbounded, but still small „small-world”

Small worlds everywhere

Social networks ( acquaintances, movie stars, Milgram 1967)

Popular models II.

Watts and Strogatz (random rewiring and deletion)
Barabási et al (growing, „preferential attachment”-idea)
Vicsek group (deterministic algorithms)

Back to the neural business

Structural questions
Constraint on locality
Different time scales

Universal tuning:the Hebbian learning (1949)

Some examples

Antisymmetric...

Prediction

Further examples

…and symmetric interaction

Synchronicity

Properties of the HebbNet II.

Properties of the HebbNet III.

Properties of the HebbNet IV.

Local mechanism
Self-organizing
Small world
Robust against noise
Ability to learn prediction and synchronization

Interpretation?

‘Neural darwinism’ theory of Edelman
Efficient memory systems
Relevant model of the Internet