Dawei W. Dong                RESEARCH STATEMENT



My research emphasizes finding fundamental theories that explain how
the nervous system codes and uses sensory information.  The approach
is based on ideas of optimal coding from information theory, for
example, decorrelation of sensory inputs to make statistically
independent representations.  The basic assumption is that such
representation of sensory signals facilitates the performance of
perception and cognition and gives an evolution advantage to the
brain.

Assuming that the brain utilizes such optimal codings, one can in
principle predict the neural processing.  Making such prediction
involves: (1) measuring the statistical properties of sensory signals
in an animal's natural environment, and then, (2) mathematically
deriving the optimal transforms for those sensory signals.  The
predicted transforms can then be compared with the actual processing
observed in physiological and psychophysical experiments.

The visual sense provides unique opportunities for highly structured
inputs, of which the statistical properties can be quantitatively
measured, and to which neural responses have been extensively studied.
Therefore, I often test my theories on the visual system.  In
particular, I test my theories on the visual systems of humans and of
other vertebrates whose visual systems share key features of
evolutionary kinship with humans and are believed to be highly
optimized for information processing.  I am now particularly
interested in the dynamic/non-stationary aspects of vision.  

In collaboration with other scientists, I am engaged in a number of
related research topics:


 * Statistics of natural images.  In order to understand how the
   visual system processes input information, one needs to understand
   more about the properties of visual inputs under natural
   conditions.  I have collected many samples of natural time-varying
   images and analyzed their spatiotemporal statistics.  This work has
   helped the currently emerging understanding of the statistics of
   spatiotemporal correlations and their relation to motions of the
   observer relative to visual scenes.  Further more, we take
   eye-movements into account and investigate their effect on the
   image statistics and other non-stationary properties of the visual
   input.

 * Optimal coding.  From the measured properties of natural
   time-varying images, the spatial-temporal coding of the lateral
   geniculate nucleus (LGN), a processing network between retina and
   cortex, can be predicted; the predictions are in quantitative
   agreement with available physiological data and have been used to
   guide new experiments.  In particular, I predict the dynamic nature
   of the LGN responses, i.e., dependence on the saccade timings and
   scenes.  I have also predicted quantitatively the perceptual shifts
   caused by adaptation to different sensory input structures, in
   particular, the orientation illusions; these predicted shifts,
   which are believed to have cortical origin, are consistent with
   recent psychophysical and physiological data.

 * Learning dynamics.  It is of great interests to study how
   neural networks self-organize to learn the desired optimal coding
   under natural conditions.  The goal here is to find convergent,
   biologically plausible development algorithms that are driven by
   neuronal inputs and outputs.  I have found one such algorithm to
   generate optimal information processing networks.  In simulation,
   when given natural time-varying images, this algorithm produces the
   spatial-temporal receptive fields of early visual pathways.
   Current simulations, with both feedforward and feedback
   connections, produce the columnar organization observed in striate
   cortex.

 * Biophysical model.  We investigate if well-known ion channel
   properties can facilitate information-theoretic optimal coding
   through temporal decorrelation; and if so, whether the degree of
   temporal decorrelation can be adapted dynamically to ensure such
   optimization at longer time scales.  We found that dynamic
   decorrelation is obtainable by varying temporal filtering through
   changing the resting membrane potential, for example, the
   biophysical properties of LGN cells support the role of temporal
   decorrelation and enable a plausible feedback control mechanism
   when input statistics change.

 * Brain imaging.  We are developing a new method of functional MRI
   study of brain activities and integrating eye-tracking into MRI
   experiments.  We compare the information processing during two
   states achieved by `free viewing' vs.  `forced viewing' of natural
   time-varying images.  During the `free viewing', a subject is shown
   a segment of a natural movie and will view it with his/her own
   natural eye movements.  In the `forced viewing' conditions the
   subject will fixate at a cross on the screen and the movie will be
   played back to the subject with the shifts in space and time
   corresponding to the subject's eye movements during one of the
   `free viewing' trials, so that the resulting retinal visual input
   will be the same.  We found different activities for the two
   conditions in both sensory visual areas and ocular motor areas.

 * Neurophysiological experiment. The single unit activities of
   LGN cells in a awake and free viewing cat are recorded during and
   between watching natural time varying images.  Simultaneously, the
   eye movement of cat is recorded.  The recorded eye positions are
   used to derive the actual visual input on the retina.  The
   auto-correlation of the the spike train and its cross-correlation
   with retinal input are calculated to reveal the STRF properties.
   We find that The spike train is temporally decorrelated not only on
   the average but also for different saccade timings and scenes.  We
   showed that both right before/after and between saccades, the
   information transfer through LGN is optimized during natural
   viewing.

 * Neurological experiment. We tracked the eye movements of
   Parkinsons Disease (PD) patients during free-viewing natural
   time-varying images.  Among the PD patients we have access to are
   those that are receiving deep-brain stimulation (DBS) to the
   subthalamic nucleus and inner sector of the globus pallidus.  We
   measured the statistics of eye movements in these individuals, with
   and without DBS, and compared them with normal controls,
   free-viewing the same natural time-varying images.  We showed that
   under natural viewing conditions, the basal ganglia play a crucial
   role in regulating and perhaps initiating the saccadic eye
   movements.  The statistics of the eye movements, including the
   fixation time and the number and the amplitude of saccades, might
   be used as sensitive indicators of PD status and the basal ganglia
   function.