Research Interests
Recent advances in signal and image processing and machine learning
algorithms have led to new methods for information processing systems. The ultimate goal of our research is to develop learning algorithms that help elucidate the details of the higher-level sensory
processes and that can be applied to practical applications for intelligent speech and image processing.
At the end of each
item, the relevant researchers
are linked for details.
Speech signal
processing for robust speech recognition
We are interested in processing speech
signals for robust speech recognition in
cars and robot toys, a hot issue in speech
recognition fields. Among the research topics
are blind deconvolution, speech denoising,
speech detection and robust feature extraction.
We effectively removed the passenger voices
or music signals using two microphones in
car navigation systems. Remaining background
noise was reduced by subsequent processing.
We evaluated the
performance of speech recognition using ICA coefficients and analyzed
speech features by using the Bayesian principal component analysis. We also
have interests in application of machine
learning algorithms to diverse problems
in speech recognition. Emotion
recognition by speech will be the next subject so that the robot
toys can respond
intelligently to the talker's emotion state. For details of blind deconvolution
and speech denoising, visit here.
For speech detection and speech feature
analysis, visit
here.
Biological and biomedical data analysis
Our research interest is in development of machine learning techniques
and application to the analysis of biological and biomedical data. We have developed the variational Bayesian learning for independent
component analysis (ICA). The ICA describes data as linear mixture of
independent features and finds projections that may uncover interesting
structure in the data. The newly developed variational Bayesian technique
allows us to work directly on datasets with high dimension but a small number
of expensive examples. The technique has been successfully applied to
a glaucoma visual field dataset to extract independent features. We have
also generalized the technique to deal with data containing missing
entries. It has been applied to a primate brain volumetric dataset and showed
very
promising results. For details, visit here.
Understanding human visual information
processing
To understand the brain we need to
answer what it does, how, and why. The way in which such a highly-organized
structure develops is also to be answered. Currently we work on the
computational modeling of the early visual system (low-level vision) and
obtained self-organizing neural network models of the LGN and V1, which is comparable
to the physiological data. We extended the modeling to a much more
realistic data set which includes spatial, temporal, and chromatic properties
with the eye movements. We are also interested in the engineering applications
such as image conversion, image compression, and the design of robot
vision. For details, visit here.
Adaptive signal processing techniques using high-order statistics for
communication systems
In
order to process information signals with high-data rate and support
high-capacity, communication receivers should have fast flexible adaptability
and improved performance. This goal can be achieved by designing efficient and
enhanced receivers with simple structures and rapid convergence characteristics.
In addition, linear/nonlinear signal processing techniques to improve the
performance of conventional linear systems are also required. This works aim to
develop efficient and robust signal processing algorithms and, simple and
enhanced linear/nonlinear receivers for high-data rate digital communication
systems.
The blind
signal separation method such as ICA is a novel
nonlinear signal processing algorithm that performs linear data transformation
that exploits high-order statistical information to project the data along the
direction of maximal independence. The framework of the
ICA is inherently analogous to that of multi-input multi-output (MIMO) systems. As
a nonlinear signal processing technique, the ICA
has demonstrated the enormous potential in many signal processing research
areas. Viewing these facts, we expect that the
ICA
can be efficiently used to suppress interference and demodulate
information data in MIMO wireless communication systems. For details, visit here.
|
