Aadvanced Research Seminar
Neural Computation and Signal
Processing Lab (NCSP)
סמינר
מתקדם במדעי
המחשב
0368-5020-01, Spring 2004-5
Prof. Nathan Intrator
Wednesday
The
seminar this semester will concentrate computational methods which are used in
the different projects in the lab. Specifically, we shall concentrate on
machine learning, robust statistics and robust modeling in each of the
presentations. The background for the specific projects has been presented in
the talks in the autumn semester
and thus, will not be repeated. There will be few guest presentations as well.
Class
Meetings: Wednesday
Instructor:
Prof. Nathan Intrator, Schreiber 221,
x7598, Office hours: Wednesday 4-5 or by appointment via email
Other
relevant seminars: PBC Seminar
Presentations
|
Date |
Title |
Speaker |
|
Feb 23 |
Organizational
meeting |
Nathan
Intrator |
|
Mar 02 |
Tal
Steinhertz |
|
|
Mar 09 |
|
|
|
Mar 23 |
||
|
Apr 13 |
DT-MRI
partial volume effects |
Ofer
Pasternak |
|
May 04 |
Niv Efron |
|
|
May 18 |
Analysis
of Doppler and Vibro-Acoustic Heart
Data |
Guy Amit |
|
May 25 |
Workshop
presentations MAX Music tool and EKG analysis |
|
|
Jun 01 |
Analysis techniques for microarray time series gene
expression data |
Omer
Berkman |
Some
reading material
|
Sound
analysis Auditory display of
hyperspectral colon tissue images Biomedical
signals and sensors Biomedical
Signal Analysis R. M. Rangayyan Breath Sounds Methodology N.
Gavriely Digital
Signal Analysis: A Computer Science Perspective J. Stein. Robust
measurement of Carotid Heart sound delay |
Software TinyOS operating sys for wireless applications Sensors Cheap off-the
shelf TinyOs operated robots PicoRadio: Low
power wireless node with sensors Machine
learning and Statistics Pattern
recognition and neural networks B. Ripley |
Abstracts
- Loop Investigation for Cursive Handwriting Tal Steinhertz
The talk will provide some
background on cursive handwriting including basic issues such as cursive script
vs. character and hand print, online vs. offline and some other principles. The
important role of loops in cursive handwriting will be described as well as the
motivation to investigate loops and learn how they were produced - in order to
support character discrimination and recognition, and contribute to writer
modeling for identification and examination (required in forensic science,
graphology etc.).
Next I will talk about the models
that I have developed to represent handwritten loops and the
features, preprocessing techniques and experts that were derived to
comply with them. I will finish my lecture with loop classification and
detection results; Recovering hidden loops on tarsi, and resolving various
properties of axial loops in letters like a, o and d at the beginning of words.
This
talk concerns with continuously searching real valued signals for a similar
segment under the Euclidean norm. his kind of continuous similarity search
(CSIMS) is a generalized similarity search (SIMS) where the space of search-set
objects is a subset of the space of queries. The currently best known algorithm
for CSIMS is the exhaustive search using normalized cross-correlation that is
computed efficiently using FFT. Currently known algorithms for SIMS such as
$kd$-tree do not work well for CSIMS due to the curse of dimensionality. An
enhancement of a few approximate SIMS algorithms for the problem of CSIMS will
be presented. A few preliminary results of an empirical evaluation of the
algorithms
will be presented.
- DT-MRI partial volume effects reduction using
the multiple tensor variational framework.
Ofer
Pasternak
Diffusion
Tensor Magnetic Resonance Imaging (DT-MRI) became a popular tool for analysis
of Diffusion Weighted magnetic resonance Images (DWIs). It provides
quantitative measures of water molecule diffusion anisotropy and the ability to
delineate and visualize major brain neuronal fiber bundles.
The
diffusion model of DT-MRI was found to be inappropriate in cases of partial
volume, where more than one type of diffusion compartment resides in the same
voxel. Partial volume could cause fiber orientation ambiguity, or
Cerebro-Spinal Fluid (CSF) contamination, both damaging the credibility and
effectiveness of fiber delineation. Diffusion models that assume higher tissue
complexity might reduce the partial volume effect, but such models have more
free parameters than DT-MRI and their fitting process for conventional
diffusion data is ill-posed. We offer the Multiple Tensor Variational (MTV)
framework as a fitting procedure which adds biologically driven constraints to
a multiple diffusion tensor model, while preserving the tensor attributes. The
multiple diffusion tensor model assumes that each voxel contains a number of
separated diffusion compartments, and the regularization terms added by the MTV
framework constrain the shape of the compartments, thus stabilizing the fitting
process. The regularized fitting allows the identification of fiber
compartments and also reduces noise by smoothing neighborhood variations. We
demonstrate how fiber ambiguity can be resolved by applying the MTV framework
on synthetic data resembling crossing neuronal fibers and on a phantom, which
simulates crossing fibers in an MRI environment similar to regular diffusion
acquisitions applied on humans. We further demonstrate how the MTV framework
can delineate larger fiber areas in cases of patients with Hydrocephalus, where
DT-MRI encounters high CSF contamination.
This talk
presents boosting methods and their application on DNA microarray
classification. We present the theory of boosting from the computational
learning perspective. We then show a statistical point of view, focusing on the
relation to additive models with different loss functions.
A brief
introduction on DNA microarrays is given, followed by 2 modified applications
of boosting on gene expression data. We cover the different methods and show
real-world and simulation results.
·
Methods for evaluating inference
algorithms Omer Berkman
A central goal of molecular
biology is to understand the regulatory mechanisms of gene transcription and
protein synthesis. This talk will review two main studies that have been done
in the field of reverse engineering of genetic networks.
A decade after the invention of DNA
Microarrays, and
after years of unsuccessful modeling of time series data, the simulation
approach has been proposed. This provides, most of all, a way to evaluate the
algorithm been used and thus ways to improve its performances.
How to build a
realistic simulator? How to sample from it? How to evaluate the interfering
algorithm? Can genetic networks be at all identified from mRNA data?
This review will give answers for this open questions and will show how much can be learned from the data in this unfavourable situation.