A Laboratory for Neural Computation Publication
In Society for Neuroscience Abstracts , 1998.
Development and Function of Spatially Invariant Feature
Hierarchies for Visual Recognition.
Gary R. Holt
and
Bartlett W. Mel
Department of Biomedical Engineering
University of Southern California
ABSTRACT
Studies of receptive fields in V4 and IT suggest that visual
object recognition is largely accomplished by a hierarchy of
feature detectors, where the features at higher levels are
increasingly form-selective but spatially invariant. This
simple, largely feedforward scheme is supported by data
suggesting that visual recognition is too fast for complicated
network interactions (e.g., Fabre-Thorpe et al.,
Neuroreport 9:303, 1998), and by the practical success of
this class of architectures on some difficult problems (e.g.,
Fukushima, IEEE Trans. Sys. Man. Cyber.
SMC-13:826, 1983; Mel, Neural Comp. 9:777, 1997).
However, the design principles underlying such a hierarchy
remain to be elucidated. What features should be present?
What degree of invariance should be introduced at each level?
What nonlinearities should be used to compute the features?
Furthermore, the means for learning the appropriate
selectivities and invariances at each level of the hieararchy
remain poorly understood. Building on work of previous
investigators such as Wallis and Rolls ( Prog.
Neurobiol. 51:167, 1997), we address with simulations (1)
how the kinds of nonlinearities present in biological neurons
can be used to best obtain invariance and selectivity, (2) how
to break down a particular recognition task into multiple
layers of detectors in order to obtain the most robust
recognition and best generalization, (3) how supervised and
unsupervised learning should interact with heuristics to guide
the search for good sets of weights in each layer in order to
solve difficult recognition problems. Supported by NIMH and
ARO and ONR.