this paper, we propose a Bayesian theory of hierarchical cortical computation based both on (a) the mathematical and computational ideas of computer vision and pattern the- ory and on (b) recent neurophysiological experimental evidence. We ,2 have proposed that Grenander's pattern theory 3 could potentially model the brain as a generafive model in such a way that feedback serves to disambiguate and 'explain away' the earlier representa- tion. The Helmholtz machine 4, 5 was an excellent step towards approximating this proposal, with feedback implementing priors. Its development, however, was rather limited, dealing only with binary images. Moreover, its feedback mechanisms were engaged only during the learning of the feedforward connections but not during perceptual inference, though the Gibbs sampling process for inference can potentially be interpreted as top-down feedback disambiguating low level representations? Rao and Ballard's predictive coding/Kalman filter model 6 did integrate generafive feedback in the perceptual inference process, but it was primarily a linear model and thus severely limited in practical utility. The data-driven Markov Chain Monte Carlo approach of Zhu and colleagues 7, 8 might be the most successful recent application of this proposal in solving real and difficult computer vision problems using generafive models, though its connection to the visual cortex has not been explored. Here, we bring in a powerful and widely applicable paradigm from artificial intelligence and computer vision to propose some new ideas about the algorithms of visual cortical process- ing and the nature of representations in the visual cortex. We will review some of our and others' neurophysiological experimental data to lend support to these ideas

There is an ongoing debate over the capabilities of hierarchical neural feed-forward architectures for performing real-world invariant object recognition. Although a variety of hierarchical models exists, appropriate supervised and unsupervised learning methods are still an issue of intense research. We propose a feedforward model for recognition that shares components like weightsharing, pooling stages, and competitive nonlinearities with earlier approaches, but focus on new methods for learning optimal featuredetecting cells in intermediate stages of the hierarchical network.

A detailed neural model is presented of how the laminar circuits of visual cortical areas V1 and V2 implement context-sensitive binding processes such as perceptual grouping and attention. The model proposes how specific laminar circuits allow the responses of visual cortical neurons to be determined not only by the stimuli within their classical receptive fields, but also to be strongly influenced by stimuli in the extra-classical surround. This context-sensitive visual processing can greatly enhance the analysis of visual scenes, especially those containing targets that are low contrast, partially occluded, or crowded by distractors. We show how interactions of feedforward, feedback and horizontal circuitry can implement several types of contextual processing simultaneously, using shared laminar circuits. In particular, we present computer simulations which suggest how top-down attention and preattentive perceptual grouping, two processes that are fundamental for visual binding, can interact, with attentional enhancement selectively propagating along groupings of both real and illusory contours, thereby showing how attention can selectively enhance object representations. These simulations also illustrate how attention may have a stronger facilitatory

Abstract—Most computer vision applications require the reliable detection of boundaries. In the presence of outliers, missing data, orientation discontinuities, and occlusion, this problem is particularly challenging. We propose to address it by complementing the tensor voting framework, which was limited to second order properties, with first order representation and voting. First order voting fields and a mechanism to vote for 3D surface and volume boundaries and curve endpoints in 3D are defined. Boundary inference is also useful for a second difficult problem in grouping, namely, automatic scale selection. We propose an algorithm that automatically infers the smallest scale that can preserve the finest details. Our algorithm then proceeds with progressively larger scales to ensure continuity where it has not been achieved. Therefore, the proposed approach does not oversmooth features or delay the handling of boundaries and discontinuities until model misfit occurs. The interaction of smooth features, boundaries, and outliers is accommodated by the unified representation, making possible the perceptual organization of data in curves, surfaces, volumes, and their boundaries simultaneously. We present results on a variety of data sets to show the efficacy of the improved formalism. Index Terms—Tensor voting, first order voting, boundary inference, discontinuities, multiscale analysis, 3D perceptual organization. 1

This paper surveys recent findings in neuroscience regarding the behavioral relevancy of the precise timing with which real spiking neurons emit spikes. The literature suggests that in almost any system where the processing-speed of a neural (sub)-system is required to be high, the timing of single spikes can be very precise and reliable. Additionally, new, more refined methods are finding precisely timed spikes where previously none where found. This line of evidence thus provides additional motivation for researching the computational properties of networks of artificial spiking neurons that compute with more precisely timed spikes.

Nonlinear contraction theory allows surprisingly simple analysis of synchronisation phenomena in distributed networks of coupled nonlinear elements. The key idea is the construction of a virtual contracting system whose particular solutions include the individual subsystems ’ states. We also study the role, in both nature and system design, of co-existing “power” leaders, to which the networks synchronize, and “knowledge” leaders, to whose parameters the networks adapt. Also described are applications to large scale computation using neural oscillators, and to time-delayed teleoperation between synchronized groups. Similarly, contraction theory can be systematically and simply extended to address classical questions in hybrid nonlinear systems. The key idea is to view the formal definition of a virtual displacement, a concept central to the theory, as describing the state transition of a differential system. This yields in turn a compositional contraction analysis of switching and resetting phenomena. Applications to hybrid nonlinear oscillators are also discussed.

Contour integration in low-level vision is believed to occur based on lateral interaction between neurons with similar orientation tuning. How such interactions could arise in the brain has been an open question. Our model suggests that the interactions can be learned through input-driven self-organization, i.e. through the same mechanism that underlies many other developmental and functional phenomena in the visual cortex. The model also shows how synchronized firing mediated by these lateral connections can represent the percept of a contour, resulting in performance similar to that of human contour integration. The model further demonstrates that contour integration performance can differ in different parts of the visual field, depending on what kinds of input distributions they receive during development. The model thus grounds an important perceptual phenomenon onto detailed neural mechanisms, so that various structural and functional properties can be measured, and predictions can be made to guide future experiments. 1

ABSTRACT. Recent findings in neuroscience strongly suggest that an object’s features (e.g., its color, texture, shape, etc.) are represented in separate areas of the visual cortex. Although represented in separate neuronal areas, somehow the feature representations are brought together as a single, unified object of visual consciousness. This raises a question of binding: how do neural activities in separate areas of the visual cortex function to produce a feature-unified object of visual consciousness? Several prominent neuroscientists have adopted neural synchrony and attention-based approaches to explain object feature binding. I argue that although neural synchrony and/or attentional mechanisms might function to disambiguate an object’s features, it is difficult to see how either of these mechanisms could fully explain the unity of an object’s features at the level of visual consciousness. After presenting a detailed critique of neural synchrony and attention-based approaches to object feature binding, I propose interactive hierarchical structuralism (IHS). This view suggests that a unified percept (i.e., a feature-unified object

This paper studies new spike-based models for winner-take-all computation and coincidence detection. In both cases, very fast convergence is achieved independent of initial conditions, and network complexity is linear in the number of inputs. Fully distributed versions can be derived by using groups of interneurons connected through electrical synapses. 1

Model building by perturbation