1. Spontaneous activity in V1: a probabilistic framework
Gergő Orbán Volen Center for Complex Systems Brandeis University
Sloan Swartz Centers Annual Meeting, 2007

2. Normative account for visual representations
Optimization criterion for the emergence of simple-cell receptive fields: independent ‘filters’ + sparseness (Bell & Sejnowski, 1995; Olshausen & Field, 1996)

3. Activity in V1 The spectrum of V1 physiology is much richer
Spontaneous activity
Response variabilty
Temporal dynamics
Can we devise a framework that
Gives a functional description of visual processing
Uses normative principles in probabilistic learning
Gives a more complete interpretation of V1 activity?

4. Computational paradigm
Density estimation
Statistically well founded
principle
Allows the representation of uncertainty
Efficient for making
predictions
Internal representation:
Useful representation
Biologically plausible
: retinal image/ RGC output;
: neural activity

5. Spontaneous activity
In the awake brain there is patterned neural activity not directly related to the stimulus
Evoked
Spontaneous
(Tsodyks et al, 1999)
Patterns of neural activities are similar in stimulus evoked condition and closed eye condition
Long-range correlations in neural activity
(Fiser et al, 2004)

6. Probabilistic model: Field of experts
Filters are componenets in a Boltzmann energy function (Osindero, Welling & Hinton, 2006)
Sparse prior (Student-t distribution)
Image model assuming translational invariance (Black & Roth, 2005)
Learning: standard contrastive divergence & Hybrid MC (Hinton 2002)
Receptive fields

7. Spontaneous activity as prior sampling
Evoked activity:
ANSATZ:
Spontaneous activity:
Evoked
activity
Natural
image
statistics
Intuitive link between evoked and spontaneous activities

8. Images generated by the model
Prior over activities
Sampling
Neural activities
Filters
Dreamed image
Images generated from prior have long-range structure

9. Evoked and spontaneous neural activity
Correlation between
hidden units
Experiment
(Fiser et al, 2004)
Evoked and spontaneous activities have similar correlational structure

10. Spontaneous neural activity before learning
Experiment
(Fiser et al, 2004)
Correlational patterns in the activity of neurons is a result of learning in the probabilistic model

11. Conclusions
The probabilistic framework provides a viable explanation for spontaneous activity in V1
Spontaneous activity as sampling from prior
Long range correlations are present both in evoked and spontaneous activities
The tendency of changes in spatial correlations with training match experimental results
Sleep-wake

12. Bottom line In the probabilistic framework:
Temporal dynamics top-down/ lateral interactions
Spontaneous activity prior sampling
Response variablity posterior variance

13. Special thanks to
Pietro Berkes (Gatsby)
Collaborators:
Máté Lengyel (Gatsby)
József Fiser (Brandeis)

16. High-level computational principles + physiology
Computational paradigm: Normative probabilistic model
Experimental paradigm: Spontaneous activity in V1
– prior sampling
– posterior variance
– top-down/
lateral interactions
Are there sensible interpretations that assign functional roles for the spontaeous activity?