We effortlessly perceive objects as having certain properties, such as shape and color, largely regardless of the conditions under which we view them. Nevertheless, the light signals entering our eyes confound information about object color, pose, shape, etc., due to the projection of a 3D world with its reflecting surfaces and illuminants onto a 2D retina with four photoreceptor types. What computations allow the brain to extract invariant information from the sensory signal? My research focuses on understanding the computational and neural mechanisms underlying perceptual constancy by employing methods from psychophysics, computational modeling, and fMRI.

I use color perception as a model system to probe this question for three primary reasons: first, a lot is known about the physiology underlying basic color perception phenomena, such as chromatic adaptation and opponent color processing; second, color stimuli are easy to control through display calibration, and third, surface color is an important source of information about object properties, such as edibility. I have recently started investigating constancy phenomena for more complex stimuli, such as 3D shapes and faces, and I find it an interesting question in its own right whether the same or analogous phenomena exist at different levels of the information processing hierarchy. Click on the links below to jump to each project description.

The role of prior knowledge in color perception/Bayesian model

Functional benefits of neural adaptation (shapes)

Top-down effects in fMRI adaptation (faces)

The role of memory in color perception/Bayesian model

Prior knowledge affects the way we interpret incoming sensory signals, both based on long-term learning (memory colors), and short-term learning (statistical priors). In my own research, I have discovered that prior knowledge about object identity affects the way we perceive their colors (see e.g Olkkonen, Hansen, & Gegenfurtner, 2008 ). More recently, I found that prior knowledge acquired on the shorter term also affects color appearance in delayed color matches ( Olkkonen, McCarthy, & Allred, 2014).

The effect of long-term or short-term memory processes on color appearance are not explained by current models of color perception or memory, but fit well in a probabilistic inference framework based on a Bayesian ideal observer. A Bayesian observer estimates the external cause of an incoming sensory signal by combining the sensory evidence with prior information about the world. Together with Toni Saarela and Sarah Allred, I have implemented a Bayesian model observer that produces similar interactions between perceptual constancy and short-term memory for lightness that we observed recently in human observers for both lightness and hue (Olkkonen & Allred, 2014 ; Olkkonen, Saarela, & Allred, in prep (OSA 2014 abstract))). My next goal is to implement this model in full-color scenes, and test the model with a new, independent data set.

Back to top

The functional benefits of neural adaptation

In another line of research, also related to invariant perception, I am investigating the functional benefits of sensory adaptation in mid-level vision. One common benefit assigned to adaptation is to help maintain sensitivity under environmental changes, achieved by shifting the response range of neuronal mechanisms in response to adaptation. This is clearly the case at the photoreceptor level, where the cells change their sensitivity according to the prevailing light levels. This allows us to see from starlight to bright sunlight although individual photoreceptors have a response range of about 100:1. Neurons after the photoreceptor level adapt also: some contrast-sensitive neurons at the retina and in early visual cortex shift their tuning curves on the contrast axis in response to contrast adaptation. This is beneficial presumably to maintain maximum sensitivity to changes in contrast regardless of the absolute contrast level.

On a more functional level, it has been suggested that adaptation serves to improve discriminability of stimuli, or to decorrelate neural signals. At the level of individual neurons, this would manifest as the sharpening of neuronal tuning curves (increased selectivity), and at the level of populations, the sharpening of population tuning curves (selectivity for the whole population). There is some evidence for the sharpening hypothesis from electrophysiology and fMRI for simple stimulus features (e.g. orientation), but little behavioral evidence for other stimuli than color (which is a salient exception), and no fMRI evidence for more complex stimuli.

Nevertheless, the sharpening hypothesis is an attractive one, and despite the lack of evidence, it is still favored by some researchers. As there are such clear benefits for color discrimination from color adaptation, and some evidence for faces, I intend to find out whether we see such benefits for object shape both from behavioral discrimination thresholds and from fMRI pattern discriminability.

In order to study the effect of adaptation on shape representations, Toni Saarela and I have developed an Octave/Matlab toolbox for generating parametric 3D radial frequency patterns, which will be available soon under an open source license. Stay tuned!

Back to top

The effects of expectation on fMRI adaptation for faces

Recent fMRI studies have suggested that fMRI adaptation in face-selective cortex is at least partly due to the formation and maintenance of top-down expectations, rather than to bottom-up effects such as synaptic depression (fatigue) (see e.g. Summerfield et al, 2008). This suggestion was based on the discovery that repetition suppression (fMRI adaptation) was modulated by how frequent the repetitions were: when repetitions were more frequent, and thus more probable, repetition suppression was stronger than when repetitions were infrequent. This result has been replicated for faces, but not for objects, casting doubt on the generality of the finding. In this project, I study whether stimulus expectations (e.g. about the range of stimulus differences, or their probability) affect fMRI adaptation for faces in a parametric face space. I presented the first findings from this project at SfN 2014 (see abstract).

Back to top