Search PubMed⌕ Search

Biomedical subjects

T Poggio

Publications and source records attributed to T Poggio.

At least 37 records · Page 2Linked to original sources
PubMed1992

Fast perceptual learning in visual hyperacuity.

In many different spatial discrimination tasks, such as in determining the sign of the offset in a vernier stimulus, the human visual system exhibits hyperacuity by evaluating spatial relations with the precision of a fraction of a photoreceptor's diameter. It is proposed that this impressive performance depends in part on a fast learning process that uses relatively few examples and that occurs at an early processing stage in the visual pathway. This hypothesis is given support by the demonstration that it is possible to synthesize, from a small number of examples of a given task, a simple network that attains the required performance level. Psychophysical experiments agree with some of the key predictions of the model. In particular, fast stimulus-specific learning is found to take place in the human visual system, and this learning does not transfer between two slightly different hyperacuity tasks.

Algorithms↗
PubMed1991

Models of object recognition.

Progress in the understanding of visual recognition in the past year has been signified by the demonstration of computational feasibility of and psychophysical support for two-dimensional view-interpolation methods.

Animals↗
PubMed1990

A network that learns to recognize three-dimensional objects.

The visual recognition of three-dimensional (3-D) objects on the basis of their shape poses at least two difficult problems. First, there is the problem of variable illumination, which can be addressed by working with relatively stable features such as intensity edges rather than the raw intensity images. Second, there is the problem of the initially unknown pose of the object relative to the viewer. In one approach to this problem, a hypothesis is first made about the viewpoint, then the appearance of a model object from such a viewpoint is computed and compared with the actual image. Such recognition schemes generally employ 3-D models of objects, but the automatic learning of 3-D models is itself a difficult problem. To address this problem in computational vision, we have developed a scheme, based on the theory of approximation of multivariate functions, that learns from a small set of perspective views a function mapping any viewpoint to a standard view. A network equivalent to this scheme will thus 'recognize' the object on which it was trained from any viewpoint.

Artificial Intelligence↗
PubMed1989

A parallel algorithm for real-time computation of optical flow.

The precise management of two-dimensional field of velocities from time-varying two-dimensional images is impossible in general. It is, however, possible to compute suitable 'optical flows' that are qualitatively similar to the velocity field in most cases. We describe a simple, parallel algorithm that computes an optical flow from sequences of real images, which is consistent with human psychophysics and suggests plausible physiological models. In particular, our algorithm runs on a Connection Machine supercomputer in close-to-real time. It shows several of the same 'illusions' that are perceived by humans. A natural physiological implementation of the model is consistent with data from cortical areas V1 and MT.

Algorithms↗
PubMed1988

Parallel integration of vision modules.

Computer algorithms have been developed for several early vision processes, such as edge detection, stereopsis, motion, texture, and color, that give separate cues to the distance from the viewer of three-dimensional surfaces, their shape, and their material properties. Not surprisingly, biological vision systems still greatly outperform computer vision programs. One of the keys to the reliability, flexibility, and robustness of biological vision systems is their ability to integrate several visual cues. A computational technique for integrating different visual cues has now been developed and implemented with encouraging results on a parallel supercomputer.

Algorithms↗
PubMed1988

Computing texture boundaries from images.

Recent computational and psychological theories of human texture vision assert that texture discrimination is based on first-order differences in geometric and luminance attributes of texture elements, called 'textons'. Significant differences in the density, orientation, size, or contrast of line segments or other small features in an image have been shown to cause immediate perception of texture boundaries. However, the psychological theories, which are based on the perception of synthetic images composed of lines and symbols, neglect two important issues. First, how can textons be computed from grey-level images of natural scenes? And second, how, exactly, can texture boundaries be found? Our analysis of these two issues has led to an algorithm that is fully implemented and which successfully detects boundaries in natural images. We propose that blobs computed by a centre-surround operator are useful as texture elements, and that a simple non-parametric statistic can be used to compare local distributions of blob attributes to locate texture boundaries. Although designed for natural images, our computation agrees with some psychophysical findings, in particular, those of Adelson and Bergen (described in the preceding article), which cast doubt on the hypothesis that line segment crossings or termination points are textons.

Algorithms↗
PubMed1985

The biophysical properties of spines as a basis for their electrical function: a comment on Kawato & Tsukahara (1983).

In a theoretical study of the passive cable properties of dendritic spines Kawato & Tsukahara (1983) claim to have proved that "the dendritic spine has no significant electrical function" (from their discussion). However, Kawato & Tsukahara restrict their analysis to current inputs to spines. Since the dimensions of spines are very small, their input resistance is expected to be very large and the synaptic input to spines has to be modeled as conductance change. Under this assumption, spines show interesting (non-linear) electrical properties: i) the somatic potential induced by an excitatory synapse on a spine may depend strongly on the shape of the spine and ii) the effect of inhibition might be confined to the spine.

Dendrites↗
PubMed1985

A simple algorithm for solving the cable equation in dendritic trees of arbitrary geometry.

We present an efficient algorithm for solving the one-dimensional cable equation in the Laplace (frequency) domain for an arbitrary linear membrane. This method, a reformulation and extension of the geometrical calculus developed by Butz and Cowan (1974), solves for the transfer impedance between any two points in a branched cable structure of arbitrary geometry (but without loops) by the repetitive application of four simple equations. Such an algorithm is used to analyze the electrical behaviour of nerve cells with highly branched dendritic trees. The algorithm can be implemented using a language such as C, PASCAL or LISP and runs on small machines.

Cell Membrane↗
PubMed1984

Vertical image registration in stereopsis.

Most computational theories of stereopsis require a registration stage prior to stereo matching to reduce the matching to a one-dimensional search. Even after registration, it is critical that the stereo matching process tolerate some degree of residual misalignment. We have studied the tolerance to vertical disparity in situations in which false targets abound and corrective eye movements are eliminated. Our main results are: vertical disparity of only the central "figure" part of a random dot stereogram can be tolerated up to about 3.5', and vertical disparity of the "figure + ground" is tolerated up to about 6.5' in the presence of monocular cues to vertical disparity. Our data suggest that this tolerance is attained by two non-motor mechanisms: the spatial average performed by the receptive fields that filter the two images prior to stereo matching, and a non-motor shift mechanism that may be driven at least in part by monocular cues.

Computers↗
PubMed1983

Nonlinear interactions in a dendritic tree: localization, timing, and role in information processing.

In a passive dendritic tree, inhibitory synaptic inputs activating ionic conductances with an equilibrium potential near the resting potential can effectively veto excitatory inputs. Analog interactions of this type can be very powerful if the inputs are appropriately timed and occur at certain locations. We examine with computer simulations the precise conditions required for strong and specific interactions in the case of a delta-like ganglion cell of the cat retina. We find some critical conditions to be that (i) the peak inhibitory conductance changes must be sufficiently large (i.e., approximately equal to 50 nS or more), (ii) inhibition must be on the direct path from the location of excitation to the soma, and (iii) the time course of excitation and inhibition must substantially overlap. Analog AND-NOT operations realized by satisfying these conditions may underlie direction selectivity in ganglion cells.

Action Potentials↗