NIPS*96 preprints

[michael@salk.edu]

Connectionists - 

This is an announcement of several NIPS*96 preprints from the
Computational Neurobiology Lab at the Salk Institute in San Diego.
These will appear in "Advances in Neural Information Processing
Systems 9" (available May 1997), edited by Mozer, M.C., Jordan, M.I.,
and Petsche, T., and published by MIT Press of Cambridge, MA.  We
enclose the abstracts and ftp addresses of these papers.  Full
citations are at the bottom of each abstract.  Comments and feedback
are welcome.

- Marni Stewart Bartlett, Tony Bell, Michael Gray, Mike Lewicki, 
	Terry Sejnowski, Magnus Stensmo, Akaysha Tang

**************************************************************

VIEWPOINT INVARIANT FACE RECOGNITION USING INDEPENDENT COMPONENT
             ANALYSIS AND ATTRACTOR NETWORKS
        Bartlett, M. Stewart & Sejnowski, T.J.

EDGES ARE THE `INDEPENDENT COMPONENTS' OF NATURAL SCENES
        Bell A.J. & Sejnowski T.J.

DYNAMIC FEATURES FOR VISUAL SPEECHREADING: A SYSTEMATIC COMPARISON 
        Gray, M.S., Movellan, J.R., & Sejnowski, T.J.

SELECTIVE INTEGRATION: A MODEL FOR DISPARITY ESTIMATION
        Gray, M.S., Pouget, A., Zemel, R., Nowlan, S., & Sejnowski, T.J. 

BLIND SEPARATION OF DELAYED AND CONVOLVED SOURCES
        Lee T-W., Bell A.J. & Lambert R.

BAYESIAN UNSUPERVISED LEARNING OF HIGHER ORDER STRUCTURE
        Lewicki, M.S. & Sejnowski, T.J.

LEARNING DECISION THEORETIC UTILITIES THROUGH REINFORCEMENT LEARNING 
        Stensmo, M. & Sejnowski, T.J.

CHOLINERGIC MODULATION PRESERVES SPIKE TIMING UNDER PHYSIOLOGICALLY
        	REALISTIC FLUCTUATING INPUT
        Tang, A.C., Bartels, A.M., & Sejnowski, T.J.

**************************************************************

VIEWPOINT INVARIANT FACE RECOGNITION USING INDEPENDENT COMPONENT
		ANALYSIS AND ATTRACTOR NETWORKS

	    Bartlett, M. Stewart & Sejnowski, T.J.

We have explored two approaches to recognizing faces across changes in
pose.  First, we developed a representation of face images based on
independent component analysis (ICA) and compared it to a principal
component analysis (PCA) representation for face recognition.  The ICA
basis vectors for this data set were more spatially local than the PCA
basis vectors and the ICA representation had greater invariance to
changes in pose.  Second, we present a model for the development of
viewpoint invariant responses to faces from visual experience in a
biological system.  The temporal continuity of natural visual
experience was incorporated into an attractor network model by Hebbian
learning following a lowpass temporal filter on unit activities.  When
combined with the temporal filter, a basic Hebbian update rule became
a generalization of Griniasty et al. (1993), which associates
temporally proximal input patterns into basins of attraction.  The
system acquired representations of faces that were largely independent
of pose.

ftp://ftp.cnl.salk.edu/pub/marni/nips96_bartlett.ps
http://www.cnl.salk.edu/~marni/publications.html

Bartlett, M. Stewart & Sejnowski, T. J. (in press).  Viewpoint
invariant face recognition using independent component analysis and
attractor networks.  In Mozer, M.C., Jordan, M.I., Petsche, T.(Eds.),
Advances in Neural Information Processing Systems 9.  MIT Press,
Cambridge, MA, U.S.A.

**************************************************************

EDGES ARE THE `INDEPENDENT COMPONENTS' OF NATURAL SCENES

              Bell A.J. & Sejnowski T.J.

Field (1994) has suggested that neurons with line and edge
selectivities found in primary visual cortex of cats and monkeys form
a sparse, distributed representation of natural scenes, and Barlow
(1989) has reasoned that such responses should emerge from an
unsupervised learning algorithm that attempts to find a factorial code
of independent visual features. We show here that non-linear
`infomax', when applied to an ensemble of natural scenes, produces
sets of visual filters that are localised and oriented.  Some of these
filters are Gabor-like and resemble those produced by the
sparseness-maximisation network of Olshausen \& Field (1996).  In
addition, the outputs of these filters are as independent as possible,
since the infomax network is able to perform Independent Components
Analysis (ICA).  We compare the resulting ICA filters and their
associated basis functions, with other decorrelating filters produced
by Principal Components Analysis (PCA) and zero-phase whitening
filters (ZCA).  The ICA filters have more sparsely distributed
(kurtotic) outputs on natural scenes.  They also resemble the
receptive fields of simple cells in visual cortex, which suggests that
these neurons form an information-theoretic co-ordinate system for
images.

ftp://ftp.cnl.salk.edu/pub/tony/edge.ps.Z

Bell A.J. & Sejnowski T.J. (In press). Edges are the `Independent
Components' of Natural Scenes. In Mozer, M.C., Jordan, M.I., Petsche,
T. (Eds.), Advances in Neural Information Processing Systems 9.  MIT
Press, Cambridge, MA, U.S.A.

**************************************************************

DYNAMIC FEATURES FOR VISUAL SPEECHREADING: A SYSTEMATIC COMPARISON 

       Gray, M. S., Movellan, J. R., & Sejnowski, T. J.

Humans use visual as well as auditory speech signals to recognize
spoken words.  A variety of systems have been investigated for
performing this task.  The main purpose of this research was to
systematically compare the performance of a range of dynamic visual
features on a speechreading task.  We have found that normalization of
images to eliminate variation due to translation, scale, and planar
rotation yielded substantial improvements in generalization
performance regardless of the visual representation used.  In
addition, the dynamic information in the difference between successive
frames yielded better performance than optical-flow based approaches,
and compression by local low-pass filtering worked surprisingly better
than global principal components analysis (PCA).  These results are
examined and possible explanations are explored.

ftp://ftp.cnl.salk.edu/pub/michael/nips_lips.ps
ftp://ftp.cnl.salk.edu/pub/michael/nips_lips-abs.text

Gray, M. S., Movellan, J. R., & Sejnowski, T. J. (In press).  Dynamic
features for visual speechreading: A systematic comparison.  In Mozer,
M.C., Jordan, M.I., Petsche, T. (Eds.), Advances in Neural Information
Processing Systems 9.  MIT Press, Cambridge, MA, U.S.A.

**************************************************************

SELECTIVE INTEGRATION: A MODEL FOR DISPARITY ESTIMATION

	Gray, M. S., Pouget, A., Zemel, R., 
	   Nowlan, S., & Sejnowski, T. J. 

Local disparity information is often sparse and noisy, which creates
two conflicting demands when estimating disparity in an image region:
the need to spatially average to get an accurate estimate, and the
problem of not averaging over discontinuities.  We have developed a
network model of disparity estimation based on disparity-selective
neurons, such as those found in the early stages of processing in
visual cortex.  The model can accurately estimate multiple disparities
in a region, which may be caused by transparency or occlusion, in real
images and random-dot stereograms.  The use of a selection mechanism
to selectively integrate reliable local disparity estimates results in
superior performance compared to standard back-propagation and
cross-correlation approaches.  In addition, the representations
learned with this selection mechanism are consistent with recent
neurophysiological results of von der Heydt, Zhou, Friedman, and
Poggio (1995) for cells in cortical visual area V2.  Combining
multi-scale biologically-plausible image processing with the power of
the mixture-of-experts learning algorithm represents a promising
approach that yields both high performance and new insights into
visual system function.

ftp://ftp.cnl.salk.edu/pub/michael/nips_stereo.ps
ftp://ftp.cnl.salk.edu/pub/michael/nips_stereo-abs.text

Gray, M. S., Pouget, A., Zemel, R., Nowlan, S., & Sejnowski, T. J. (In
press).  Selective Integration: A Model for Disparity Estimation.  In
Mozer, M.C., Jordan, M.I., Petsche, T. (Eds.), Advances in Neural
Information Processing Systems 9.  MIT Press, Cambridge, MA, U.S.A.

**************************************************************

BLIND SEPARATION OF DELAYED AND CONVOLVED SOURCES
	
	Lee T-W., Bell A.J. & Lambert R.

We address the difficult problem of separating multiple speakers with
multiple microphones in a real room. We combine the work of Torkkola
and Amari, Cichocki and Yang, to give Natural Gradient information
maximisation rules for recurrent (IIR) networks, blindly adjusting
delays, separating and deconvolving mixed signals. While they work
well on simulated data, these rules fail in real rooms which usually
involve non-minimum phase transfer functions, not-invertible using
stable IIR filters. An approach that sidesteps this problem is to
perform infomax on a feedforward architecture in the frequency domain
(Lambert 1996). We demonstrate real-room separation of two natural
signals using this approach.

ftp://ftp.cnl.salk.edu/pub/tony/twfinal.ps.Z

Lee T-W., Bell A.J. & Lambert R. (In press). Blind separation of
delayed and convolved sources. In Mozer, M.C., Jordan, M.I., Petsche,
T. (Eds.), Advances in Neural Information Processing Systems 9.  MIT
Press, Cambridge, MA, U.S.A.

**************************************************************

BAYESIAN UNSUPERVISED LEARNING OF HIGHER ORDER STRUCTURE

	    Lewicki, M. S. & Sejnowski, T. J.

Multilayer architectures such as those used in Bayesian belief
networks and Helmholtz machines provide a powerful framework for
representing and learning higher order statistical relations among
inputs.  Because exact probability calculations with these models are
often intractable, there is much interest in finding approximate
algorithms. We present an algorithm that efficiently discovers higher
order structure using EM and Gibbs sampling. The model can be
interpreted as a stochastic recurrent network in which ambiguity in
lower-level states is resolved through feedback from higher levels. We
demonstrate the performance of the algorithm on benchmark problems.

ftp://ftp.cnl.salk.edu/pub/lewicki/nips96.ps.Z
ftp://ftp.cnl.salk.edu/pub/lewicki/nips96-abs.text

Lewicki, M.S. and Sejnowski, T.J. (In press). Bayesian unsupervised
learning of higher order structure.  In Mozer, M.C., Jordan, M.I., and
Petsche, T. (Eds.), Advances in Neural and Information Processing
Systems 9.  MIT Press, Cambridge, MA, U.S.A.

**************************************************************

LEARNING DECISION THEORETIC UTILITIES THROUGH REINFORCEMENT LEARNING 

		Stensmo, M. & Sejnowski, T. J.

Probability models can be used to predict outcomes and compensate for
missing data, but even a perfect model cannot be used to make
decisions unless the values of the outcomes, or preferences between
them, are also provided. This arises in many real-world problems, such
as medical diagnosis, where the cost of the test as well as the
expected improvement in the outcome must be considered. Relatively
little work has been done on learning the utilities of outcomes for
optimal decision making. In this paper, we show how
temporal-difference (TD($\lambda$)) reinforcement learning can be used
to determine decision theoretic utilities within the context of a
mixture model and apply this new approach to a problem in medical
diagnosis. TD($\lambda$) learning reduces the number of tests that
have to be done to achieve the same level of performance with the
probability model alone, which result in significant cost savings and
increased efficiency.

http://www.cs.berkeley.edu/~magnus/papers/nips96.ps.Z

Stensmo, M. and Sejnowski, T. J. (in press). Learning decision
theoretic utilities through reinforcement learning. In: Mozer, M.C.,
Jordan, M.I. and Petsche, T., (Eds.), Advances in Neural Information
Processing Systems, Vol. 9. MIT Press, Cambridge, MA, U.S.A.

**************************************************************

CHOLINERGIC MODULATION PRESERVES SPIKE TIMING UNDER PHYSIOLOGICALLY
		REALISTIC FLUCTUATING INPUT

	Tang, A. C., Bartels, A. M., & Sejnowski, T. J.

Neuromodulation can change not only the mean firing rate of a neuron,
but also its pattern of firing.  Therefore, a reliable neural coding
scheme, whether a rate coding or a spike time based coding, must be
robust in a dynamic neuromodulatory environment.  The common
observation that cholinergic modulation leads to a reduction in spike
frequency adaptation implies a modification of spike timing, which
would make a neural code based on precise spike timing difficult to
maintain.  In this paper, the effects of cholinergic modulation were
studied to test the hypothesis that precise spike timing can serve as
a reliable neural code.  Using the whole cell patch-clamp technique in
rat neocortical slice preparation and compartmental modeling
techniques, we show that cholinergic modulation, surprisingly,
preserved spike timing in response to a fluctuating inputs that
resembles {\em in vivo} conditions.  This result suggests that in vivo
spike timing may be much more resistant to changes in neuromodulator
concentrations than previous physiological studies have implied.

ftp://ftp.cnl.salk.edu/pub/tang/ach_timing.ps.gz
ftp://ftp.cnl.salk.edu/pub/tang/ach_timing_abs.txt

Akaysha C. Tang, Andreas M. Bartels, and Terrence J Sejnowski. (In
press). Cholinergic Modulation Preserves Spike Timing Under
Physiologically Realistic Fluctuating Input.  In Mozer, M.C., Jordan,
M.I., Petsche, T. (Eds.), Advances in Neural Information Processing
Systems 9.  MIT Press, Cambridge, MA, U.S.A.

**************************************************************