Laboratory of Psychophysics LPSY
|Research topics||Curriculum Vitae||Publications||Abstracts|
What is an object? How are objects coded in the brain?
How are features of an object bound together? How long does object
binding take? How to code objects mathematically? Why are objects so
easily perceived but their physiological, psychophysical, and
computational status so hard to grasp? At the end, are there objects at
all or only features?
How can we investigate these topics? Studying the changes and dynamics of object presentation might be one approach, e.g. investigating perceptual learning. Studying the temporal processing of spatial features might be another way. Mathematical models will help to understand the interplay of the essentials of object processing. But, what are those?
Παντα ρει: Perceptual learning
In most models of perceptual learning,
learning depends on changes of synaptic weights determined by the
stimuli presented only. Quite to the contrary, we proposed that
perceptual learning is primarily guided by top-down operations (Herzog
& Fahle, 1998).
Contrary to supervised learning models, external feedback is not used to classify stimuli ("this is a banana'') but serves as a performance measure ("how much did I improve using a given top-down operation?''; Herzog & Fahle, 1997; 1998).
How plastic is the human brain? Can we learn to perceive a left tilted line to be tilted to the right if a teacher tells us so? This is not the case. External feedback biases decisions but not perception: we cannot perceive a cat as a dog but we might be willing to say so (Herzog & Fahle, 1999; Herzog, Ewald, Hermens & Fahle; 2006).
Current research focuses on the question why perceptual learning can be abolished in so called roving paradigms, i.e. when various stimuli are presented randomly interleaved, but not when presented in separate sessions (Zhaoping, Herzog & Dayan, 2003; Otto, Herzog, Fahle & Zhaoping, 2006; Parkosadze, Otto, Malania, Kezeli & Herzog, 2008).
Freeing features: Feature inheritance and non-retinotopic feature attribution
Usually, visual features of an object are
perceived at the same position as the object they belong to. A yellow
banana, for example, is perceived as yellow even when neighboring a red
apple. Nevertheless, we could show with a variety of paradigms that
low-level features can be "freed" from their physical carriers and can
be attributed to other elements. For example, a subjectively invisible
that precedes the presentation of a vertical line, surprisingly, makes
the vertical line appearing to be tilted (Herzog & Koch, 2001).
Whereas such kinds of illusory conjunctions are usually considered to
reflect limitations of the visual system, we showed that these "binding
errors" reveal a computational strategy of the human brain to cope with
the vast amount of information entering the eye. The integration of
features of elements is in accordance with the spatio-temporal grouping
of these elements (Ogmen, Otto & Herzog, 2006; Otto, Ogmen &
The importance of spatial processing is also evident in other feature integration paradigms such as feature fusion (Herzog, Parish, Koch & Fahle, 2003; Herzog, Lesemann & Eurich, 2006; Scharnowski, Hermens & Herzog, 2007; Scharnowski, Hermens, Kammer, Ogmen & Herzog, 2007; Herzog, Scharnowski & Hermens, 2007).
More is better: Visual masking and the dynamics of visual information processing
In visual masking, a mask impairs
performance on a preceding target. Traditionally, the energy ratio of
the target and mask was thought to be the most important aspect of
masking strength: The brighter a light, the more you will be blinded by
it. We showed theoretically and experimentally that energy per se often
fails to explain masking strength (Francis & Herzog, 2004). For
example experimentally it was demonstrated that masks with more
elements (higher energy) can yield weaker masking than masks with fewer
elements (lower energy). We proposed that the spatial layout of the
target and the mask is as crucial as the energy to explain masking
strength (Herzog & Koch, 2001; Herzog, Fahle & Koch, 2001;
Herzog, Koch & Fahle, 2001a,b; Herzog & Fahle, 2002; Herzog,
Harms, Ernst, Eurich, Mahmud & Fahle, 2003; Duangudom, Francis
& Herzog, 2007; Hermens & Herzog, 2007; Saarela & Herzog,
On a description level of perceptual organization, grouping seems to be most important. On a description level of neural mechanisms, a simple neural network model with dynamical lateral inhibition and excitation can explain many of these and other masking results (Herzog, Ernst, Etzold & Eurich, 2003; Hermens, Luksys, Gerstner, Herzog & Ernst, 2008; for a popular description, see Herzog, Ernst & Eurich, 2004). In general, models of temporal vision need to incorporate spatial components and models of spatial vision temporal components (Herzog, 2007).
Interestingly, masking can be strongly modulated by benzodiazepines modulating the GabaA receptor (Giersch & Herzog, 2003) and transcranial magnetic stimulation (Kammer, Scharnowski & Herzog, 2003).
Good and bad neighbors: Contextual modulation
The processing of a target stimulus is strongly influenced by neighboring stimuli. Most models of contextual modulation propose local interactions in the processing of the target and its neighbors, such as lateral inhibition or spatial pooling of information from nearby locations. Contrary to this models, we have shown that the global spatial layout of the neighboring elements strongly modulates performance. This contextual modulation can be best described in terms of perceptual grouping (temporal masking: Herzog & Fahle, 2002; Herzog, Dependahl, Schmonsees & Fahle, 2003; Herzog, Schmonsees & Fahle, 2003a,b; Hermens & Herzog, 2007; Saarela & Herzog, 2008; simultaneous masking: Malania, Herzog & Westheimer, 2007). We believe contextual modulation is at the very heart of neuroscience because it shows that perceptual responses and neural firing rates are not determined by a target stimulus alone. Rather, the global aspects of the whole stimulus have a strong influence on the processing, and thus on our perception.
To be or not to be: Conscious and unconscious processing
Visual masking is a versatile tool to
render stimuli unconscious. Often it is assumed that the whole stimulus
is suppressed by masking. We showed that features of an invisible
target can still be perceived at others elements (Herzog & Koch,
2001; Sharikadze, Fahle & Herzog, 2005; Otto, Ogmen & Herzog,
Many models have been proposed to explain consciousness employing, for example, recurrent neural connections, perception-action loops, and complexity measures. We proposed that these components are possibly necessary but not sufficient to explain consciousness. All these models suffer from an argument, we called the small network argument: for each of the proposed mechanisms, we can construct a very small network, consisting of less then 10 neurons, that is hardly conscious (Herzog, Esfeld & Gerstner, 2007). Hence, proposed mechanisms are not sufficient to explain consciousness.
Schizophrenic patients show performance deficits over a wide range. Lower processing deficits are of particular interest because they might cause higher, cognitive deficiencies. Low level visual masking has been proven to be one of the most promising paradigms for diagnostic purposes (e.g. the endophenotype concept) and to understand the underlying neural deficits. We showed that patients show strong masking deficits, i.e. they need longer SOAs to process a target. Paradoxically, their spatial and temporal resolution seems to be fast and intact (Herzog, Kopmann & Brand, 2004; Brand, Kopmann & Herzog, 2004; Brand, Kopmann, Marbach, Heinze & Herzog, 2005; Schutze, Bongard, Marbach, Brand & Herzog, 2007; Roinishvili, Chkonia, Brand & Herzog, 2008).
To ski or not to ski is not a question and the best ski has a low flexural and a high torsional rigidity (Fischer, Overney, Fauve, Blanke, Rhyner, Herzog, Bourban & Manson, 2007).