Abstract
Context exerts a dramatic influence on neural processing and sensory experience. Perceptually, the presence
of contextual information at a given point can elicit striking misjudgements of local features, such as orientation
and motion. This is manifest in illusions and aftereffects, which have been a topic of intensive study
for decades. Perceptual illusions are most puzzling when contexts induce distortions that appear inconsistent
with their statistically normative implications, as their functional role then is quite mysterious.
Here we focus on one of the best studied contextual effects, namely the influence of spatial surround on
local orientation misjudgements, i.e., the tilt illusion. We build computational models that treat populations
of orientation-tuned neurons in primary visual cortex in statistical terms, and link the outputs of such
populations with the perceptual phenomena they appear to underpin.
Specifically, we consider neural-level models of divisive gain control. This idea has a rich mechanistic and
functional pedigree. We formulate a generative version of divisive gain control as part of a well-found model
of natural image statistics (so-called Gaussian Scale Mixture Models). We then ask how this model leads
to changes in tuning curves, and consequently, through population decoding, to misjudgements in the tilt
illusion.
Previous work has shown that in natural scenes, filters with similar orientation preferences that represent
nearby locations in the image have strong statistical dependencies, and so are, correctly, members of the
same divisive gain control pool. We demonstrate that through population decoding, this contextual normalization
leads to tilt repulsion. However, an interesting consequence of contextual effects, as in the tilt
illusion, is that the nature of the misjudgements can be either repulsive or attractive, depending on factors
such as the relative angle between the center target and surrounding context stimuli. Although the repulsive
effect has been most widely modeled, the (weaker but consistently present) attractive effect is also diagnostic
of the system behavior. We formulate a variant of the model that obtains both attraction and repulsion. The
modified model assumes a more sophisticated scheme, in which the target and spatial context filters have a
probability of belonging to the same gain control pool.
Although tilt and spatial context is a particularly convenient example for which there are many diverse data,
most of the underlying issues extend to other visual attributes and contextual phenomena.