1 Bayesian inference & visual processing in the brainNeuroinformatics Group: Aapo Hyvärinen Patrik Hoyer Jarmo Hurri Mika Inki Urs Köster Jussi Lindgren Jukka Perkiö Ilmari Kurki

2 Paradoxes in perceptionPerception seems - effortless - straightforward - objective

3 Paradoxes in perceptionPerception seems - effortless - straightforward - objective In reality - it cannot be easily programmed in a computer - it seems to require complicated processing - it can be fooled

4 Example: Illusory motion

5 Example: Illusory motion

6 Example 2: completion

7 Example 2: completion = +

8 Example 2: completion (not:) NOT: = = +

9 Example 3: illusory contours

10 Visual processing as inferenceDominant school in vision research: constructivism Perception is unconscious inference Combine Hidden assumptions (priors) given by internal models Incoming sensory information to reach conclusions about the environment. (Helmholtz, late 19th century) Formalized as Bayesian inference

11 Our approach: Linear models of natural images= S + S S 1 2 N What are the best linear features for natural images?

12 Independent component analysisLinear mixtures of source signals: can we find the original ones?

13 Independent component analysis

14 Independent component analysis of natural imagesLow-level statistical prior Similar to what is found in the visual brain areas