Expertise:
Research in the Amit laboratory is concerned with understanding
how computer vision is to develop algorithms that can learn
representations of objects from training sets and subsequently
label digital images with instances of these objects. The
main focus of my research is tackling this challenge through
proper statistical modeling of object appearance. Although
not extensively used in computer vision these emerge as a
powerful tool in developing recognition algorithms which
allow for proper modeling of object and data variability.
The simplicity and transparency of the statistical models
enables training with small samples, and give rise to efficient
computational methods. Models for individual objects can
be composed to create models for entire scenes. The models
have been implemented in concrete applications such as reading
license plates on photos of cars, reading handwritten zipcodes,
detecting faces, cars or other objects in images. Similar
ideas have also been applied to the problem of speech recognition.
Although the speech recognition is a more mature field of
research than vision, there are some interesting insights
from vision that may contribute to increase robustness and
stability of speech recognition algorithms.
I am particularly interested in the relation between computer
algorithms for vision and biological processing in the cortex.
Can the computer algorithms be implemented in a biologically
plausible neural network and can they contribute to generating
hypotheses on how the visual cortex processes input. Motivated
by the vision algorithms we have developed biologically plausible
neural network models for translation invariant detection
of objects and for classification among object classes that
have been applied to real problems such as face detection
or handwritten digit recognition. The networks learn using
Hebbian synaptic modification rules. The two networks have
been integrated in a comprehensive architecture with interesting
analogies to many of the components of the biological visual
system, and the ability to explain many psychophysical and
electro-physiological experiments on recognition and attention.
The models have also lead to some interesting hypotheses
on effects of attention on responses of lower level input
neurons.
Specific research projects:
-- Efficient modeling and tracking of 3d objects
-- Simultaneous modeling of large numbers of object classes
in scenes with background clutter
-- Psychophysical and electro-physiological testing of
the hypotheses generated by the vision architecture
--Designing neural networks to compute refined instantiation
of objects
|
