Research theme: Theory of computation in human visual information processing
To understand the brain we should be able to explain what it does, how, and why
(in other words, "what it computes, how it computes, and what it should compute" --
M. Carandini).
To this end, I'm currently working on low-level vision - the early stages of the hierarchical visual system.
(Here is our lab's projects.)
A. Robust coding for the limited neural capacity
Individual neuron has a limited information capacity due to intrinsic neural noise.
This poses a question on the simplistic models of efficient sensory coding
such as sparse coding and independent component analysis
(ICA)
because redundancy should also be utilized in such a condition.
We proposed a robust coding model that yield the optimal linear filters
with limited coding precision.
This is significantly robust to noise compared with
the conventional linear image coding methods such as ICA, PCA, and wavelet (Fig. A-1).
One important feature of this model is that
redundancy can be introduced in the representation, instead of solely reduced.
Particularly, when neural noise is sufficiently large,
robust coding yields a degenerate code,
which represents identical, and therefore completely redundant, information (Fig. A-2).
Furthermore, robust coding can employ an arbitrary number of neurons
and find the optimal filters for the population.
For instance, we can make the reconstruction error arbitrarily small
by (minimally) increasing the number of coding units.
When it is applied to natural images with the additional constraint of sparse representations,
the optimal filters becomes a set of Gabor-like filters with a broader spatial-frequency tuning.
These results are in closer agreement with the physiological data
than the earlier models,
offering a new perspective on the expansion of the number of neurons from retina to V1,
and providng a theoretical model of incorporating useful redundancy into efficient neural representations
(Doi and Lewicki, 2004; Doi, Balcan, and Lewicki, 2005).
B. Robust coding with sensory noise
We have further generalized the robust coding in which
the observed signal (photoreceptor response) is also noisy.
As the noise sources we took into account
both blurring by the eye's optics and photon transduction noise,
and we derived the optimal code
that compensates for optical blur, photon transduction noise, and neural noise, simultaneouosly.
The resulting linear filters show the adaptive change
to the illumination level as in the experimental measurements (Figure below):
from center-surround type to center-only type as the the illumination level
(and hence the SNR of the signal)
decreases.
This model provides a unified theoretical account of the retinal population coding
(Doi and Lewicki, 2006).
C. Principled approach to color vision
In our previous study
(Doi et al, 2003; Wachtler et al., 2007; Doi et al., in preparation),
we investigated the similarities between
the statistical analysis methods, namely, whitening and ICA,
and the computations in the thalamus (LGN) and in the primary visual cortex (V1).
Specifically, we considered pricise anatomical and physiological data
such as
spectral sensitivities of cone photoreceptors and their arrangement on the retina (Fig. 1).
A diagram of our model is shown in Fig. 2.
Our results showed that
spatio-chromatic receptive fields of V1 cells (both simple-cell and double-opponent-cell type)
and the subjective quantities created in the brain
such as the luminosity function and the perceptual color space could emerge
solely through the interaction between the simple learning rule and the visual experience (Fig. 3).
Journal paper
Wachtler, T., Doi, E., Lee, T.-W., & Sejnowski, T. J.
Cone selectivity derived from the responses of the retinal cone mosaic to natural scnens.
Journal of Vision, accepted.
[abstract]
Doi, E., Balcan, D. C., & Lewicki, M. S.
Robust coding over noisy overcomplete channels.
IEEE Transactions on image processing, 2007, vol.16, pp.442-452.
[abstract]
[preprint]
Doi, E., Inui, T., Lee, T.-W., Wachtler, T., & Sejnowski, T. J.
Spatiochromatic receptive field properties derived from information-theoretic analyses of cone mosaic responses to natural scenes. Neural Computation, 2003, vol.15, pp.397-417.
[abstract]
[reprint]
Review article
Doi, E. & Lewicki, M. S.
Relations between the statistical regularities of natural images and the response properties of the early visual system.
Japanese Cognitive Science Society, SIG P&P. pp.1-8. Kyoto, Japan. July 2005.
[abstract]
[preprint]
Selected presentation
Doi, E. & Lewicki, M. S.
A theory of retinal population coding.
To appear in Neural Information Processing Systems Conference.
Vancouver, Canada, December 2006.
[abstract]
[preprint]
Doi, E.
"Retinal information processing" and "Mutual information or mean squared error?"
Neural Information Processing Systems Workshop in Decoding the Neural Code.
Whistler, Canada. December 2006.
Doi, E.
A top-down account for the retinal coding.
Pitt-CMU CNBC Retreat.
Pittsburgh, October 2006.
[slides]
Doi, E., Balcan, D. C., & Lewicki, M. S.
Optimal filters under biological constraints predict population coding of retinal ganglion cells.
Society for Neuroscience Annual meeting,
Atlanta, GA. October 2006.
[abstract]
Doi, E. & Lewicki, M. S.
Population coding of natural images with sensory and channel noise.
Computational and Systems Neuroscience,
Salt Lake City. March 2006.
Doi, E., Balcan, D. C., & Lewicki, M. S.
A theoretical analysis of robust coding over noisy overcomplete channels. Neural Information Processing Systems Conference, Vancouver, Canada, December 2005.
[abstract]
[preprint]
Doi, E. & Lewicki, M. S.
Sparse coding of natural images using an overcomplete set of limited capacity units. Neural Information Processing Systems Conference, Vancouver, Canada, December 2004.
[abstract]
[preprint]
[poster]
Doi, E. & Lewicki, M. S.
Robust coding over noisy channels with sparse overcomplete representations. In Overcomplete Representations, Neural Information Processing Systems Workshop, Whistler, Canada, December 2004.
[slides]
Doi, E.
A study of computational neural network models on spatio-chromatic properties of the early visual system. Dissertation. Kyoto University, Japan. March 2003.
Doi, E., Inui, T., Lee, T.-W., Wachtler, T., & Sejnowski, T. J.
Spatial and chromatic filters derived from an information-theoretic analysis of natural scenes. Society for Neuroscience Annual Meeting. San Diego. November 2001.
[abstract]
[poster]
Doi, E. & Inui, T.
Whitening and independent component filters of simulated cone mosaic output have spatio-chromatic properties of pLGN and V1. 7th International Conference on Neural Information Processing, pp.500-505. Taejon, Korea. Novemver 2000.
[abstract]
[preprint]
Doi, E. & Inui, T.
Self-organization of spatio-chromatic receptive fields in the early visual system by ICA. Technical report of IEICE, NC98-170, pp.131-138. Tokyo, Japan. March 1999.
[abstract]
[preprint]
Gabor-fit code (developed with Mike Lewicki)
A MATLAB code to fit two-dimensional data with Gabor function.
It can almost perfectly recover the original parameters
with which the two-dimensional data are generated.
[download page]
Contact Information
Dr. Eizaburo Doi
4 Washington Place, Room 809
New York, NY 10003
phone: 212-992-8752 (office)
412-559-2019 (cell)
office: Room 1027
email: edoi at cns dot nyu dot edu
www: http://www.cns.nyu.edu/~edoi
