Coding of multivariate stimuli and contextual interactions in the visual cortex
View/ Open
Date
02/07/2018Author
Keemink, Sander Wessel
Metadata
Abstract
The primary visual cortex (V1) has long been considered the main low level visual
analysis area of the brain. The classical view is of a feedfoward system functioning
as an edge detector, in which each cell has a receptive field (RF) and a preferred orientation.
Whilst intuitive, this view is not the whole story. Although stimuli outside
a neuron’s RF do not result in an increased response by themselves, they do modulate
a neuron’s response to what’s inside its RF. We will refer to such extra-RF effects
as contextual modulation. Contextual modulation is thought to underlie several
perceptual phenomena, such as various orientation illusions and saliency of specific
features (such as a contour or differing element). This gives a view of V1 as more
than a collection of edge detectors, with neurons collectively extracting information
beyond their RFs. However, many of the accounts linking psychophysics and physiology
explain only a small subset of the illusions and saliency effects: we would
like to find a common principle. So first, we assume the contextual modulations experienced
by V1 neurons is determined by the elastica model, which describes the
shape of the smoothest curve between two points. This single assumption gives rise
to a wide range of known contextual modulation and psychophysical effects. Next,
we consider the more general problem of encoding and decoding multi-variate stimuli
(such as center surround gratings) in neurons, and how well the stimuli can be decoded
under substantial noise levels with a maximum likelihood decoder. Although the maximum
likelihood decoder is widely considered optimal and unbiased in the limit of no
noise, under higher noise levels it is poorly understood. We show how higher noise
levels lead to highly complex decoding distributions even for simple encoding models,
which provides several psychophysical predictions. We next incorporate more updated
experimental knowledge of contextual modulations. Perhaps the most common form of
contextual modulations is center surround modulation. Here, the response to a center
grating in the RF is modulated by the presence of a surrounding grating (the surround).
Classically this modulation is considered strongest when the surround is aligned with
the preferred orientation, but several studies have shown how many neurons instead
experience strongest modulation whenever center and surround are aligned. We show
how the latter type of modulation gives rise to stronger saliency effects and unbiased
encoding of the center. Finally, we take an experimental perspective. Recently, both
the presence and the underlying mechanisms of contextual modulations has been increasingly
studied in mice using calcium imaging. However, cell signals extracted
with calcium imaging are often highly contaminated by other sources. As contextual
effects beyond center surround modulation can be subtle, a method is needed to remove
the contamination. We present an analysis toolbox to de-contaminate calcium
signals with blind source separation. This thesis thus expands our understanding of
contextual modulation, predicts several new experimental results, and presents a toolbox
to extract signals from calcium imaging data which should allow for more in depth
studies of contextual modulation.