Unsupervised segmentation with dynamical units

In this paper, we present a novel network to separate mixtures of inputs that have been previously learned. A significant capability of the network is that it segments the components of each input object that most contribute to its classification. The network consists of amplitude-phase units that can synchronize their dynamics, so that separation is determined by the amplitude of units in an output layer, and segmentation by phase similarity between input and output layer units. Learning is unsupervised and based on a Hebbian update, and the architecture is very simple. Moreover, efficient segmentation can be achieved even when there is considerable superposition of the inputs. The network dynamics are derived from an objective function that rewards sparse coding in the generalized amplitude-phase variables. We argue that this objective function can provide a possible formal interpretation of the binding problem and that the implementation of the network architecture and dynamics is biologically plausible.

Object segmentation and recovery via neural oscillators implementing the similarity and prior knowledge gestalt rules

Object recognition requires the solution of the binding and segmentation problems, i.e., grouping different features to achieve a coherent representation. Synchronization of neural activity in the gamma-band, associated with gestalt perception, has often been proposed as a putative mechanism to solve these problems, not only as to low-level processing, but also in higher cortical functions. In the present work, a network of Wilson-Cowan oscillators is used to segment simultaneous objects, and recover an object from partial or corrupted information, by implementing two gestalt rules: similarity and prior knowledge. The network consists of H different areas, each devoted to representation of a particular feature of the object, according to a topological organization. The similarity law is realized via lateral intra-area connections, arranged as a "Mexican-hat". Prior knowledge is realized via inter-area connections, which link properties belonging to a previously memorized object. A global inhibitor allows segmentation of several objects avoiding interference. Simulation results, performed using three simultaneous input objects, show that the network is able to detect an object even in difficult conditions (i.e., when some features are absent or shifted with respect to the original one). Moreover, the trade-off between sensitivity (capacity to detect true positives) and specificity (capacity to reject false positives) can be controlled acting on the extension of lateral synapses (i.e., on the level of accepted similarity). Finally, the network can also deal with correlated objects, i.e., objects which have some common features. Simulations performed using a different number of objects (2, 3, 4 or 5) suggest that the network is able to segment and recall up to four objects, but the oscillation frequency must increase, the lower the number of objects simultaneously present. The model, although quite simpler compared with neurophysiology, may represent a theoretical framework for the analysis of the relationships between object representation, memory, learning, and gamma-band activity. In particular, it extends previous studies on autoassociative memory since it exploits not only oscillatory dynamics, but also a topological organization of features.

The time dimension for scene analysis

A fundamental issue in neural computation is the binding problem, which refers to how sensory elements in a scene organize into perceived objects, or percepts. The issue of binding is hotly debated in recent years in neuroscience and related communities. Much of the debate, however, gives little attention to computational considerations. This review intends to elucidate the computational issues that bear directly on the binding issue. The review starts with two problems considered by Rosenblatt to be the most challenging to the development of perceptron theory more than 40 years ago, and argues that the main challenge is the figure-ground separation problem, which is intrinsically related to the binding problem. The theme of the review is that the time dimension is essential for systematically attacking Rosenblatt's challenge. The temporal correlation theory as well as its special form--oscillatory correlation theory-is discussed as an adequate representation theory to address the binding problem. Recent advances in understanding oscillatory dynamics are reviewed, and these advances have overcome key computational obstacles for the development of the oscillatory correlation theory. We survey a variety of studies that address the scene analysis problem. The results of these studies have substantially advanced the capability of neural networks for figure-ground separation. A number of issues regarding oscillatory correlation are considered and clarified. Finally, the time dimension is argued to be necessary for versatile computing.

Object segmentation and reconstruction via an oscillatory neural network: interaction among learning, memory, topological organization and gamma-band synchronization

Synchronization of neuronal activity in the gamma-band has been shown to play an important role in higher cognitive functions, by grouping together the necessary information in different cortical areas to achieve a coherent perception. In the present work, we used a neural network of Wilson-Cowan oscillators to analyze the problem of binding and segmentation of high-level objects. Binding is achieved by implementing in the network the similarity and prior knowledge Gestalt rules. Similarity law is realized via topological maps within the network. Prior knowledge originates by means of a Hebbian rule of synaptic change; objects are memorized in the network with different strengths. Segmentation is realized via a global inhibitor which allows desynchronisation among multiple objects avoiding interference. Simulation results performed with a 40x40 neural grid, using three simultaneous input objects, show that the network is able to recognize and segment objects in several different conditions (different degrees of incompleteness or distortion of input patterns), exhibiting the higher reconstruction performances the higher the strength of object memory. The presented model represents an integrated approach for investigating the relationships among learning, memory, topological organization and gamma-band synchronization.

Oscillatory model of attention-guided object selection and novelty detection

We develop a new oscillatory model that combines consecutive selection of objects and discrimination between new and familiar objects. The model works with visual information and fulfils the following operations: (1) separation of different objects according to their spatial connectivity; (2) consecutive selection of objects located in the visual field into the attention focus; (3) extraction of features; (4) representation of objects in working memory; (5) novelty detection of objects. The functioning of the model is based on two main principles: the synchronization of oscillators through phase-locking and resonant increase of the amplitudes of oscillators if they work in-phase with other oscillators. The results of computer simulation of the model are described for visual stimuli representing printed words.

Pulse-Coupled Neural Networks for Contour and Motion Matchings

Two neural networks based on temporal coding are proposed in this paper to perform contour and motion matchings. Both of the proposed networks are three-dimensional (3-D) pulse-coupled neural networks (PCNNs). They are composed of simplified Eckhorn neurons and mimic the structure of the primary visual cortex. The PCNN for contour matching can segment from the background the object with a particular contour, which has been stored as prior knowledge and controls the network activity in the form of spike series; The PCNN for motion matching not only detects the motion in the visual field, but also extracts the object moving in an arbitrarily specified direction. The basic idea of these two models is to encode information into the timing of spikes and later to decode this information through coincidence detectors and synapse delays to realize the knowledge-controlled object matchings. The simulation results demonstrate that the temporal coding and the decoding mechanisms are powerful enough to perform the contour and motion matchings.