The Specialized Mappings Architecture

A probabilistic, nonlinear supervised learning model is proposed: the Specialized Mappings Architecture (SMA). The SMA employs a set of several forward mapping functions that are estimated automatically from training data. Each specialized function maps certain domains of the input space (e.g., image features) onto the output space (e.g., articulated body parameters). The SMA can model ambiguous, one-to-many mappings that may yield multiple valid output hypotheses. Once learned, the mapping functions generate a set of output hypotheses for a given input via a statistical inference procedure. The SMA inference procedure incorporates an inverse mapping or feedback function in evaluating the likelihood of each of the hypothesis. Possible feedback functions include computer graphics rendering routines that can generate images for given hypotheses. The SMA employs a variant of the Expectation-Maximization algorithm for simultaneous learning of the specialized domains along with the mapping functions, and approximate strategies for inference. The framework is demonstrated in a computer vision system that can estimate the articulated pose parameters of a human’s body or hands, given silhouettes from a single image. The accuracy and stability of the SMA are also tested using synthetic images of human bodies and hands, where ground truth is known.

Specialized Mappings Architecture with Applications to Vision-based Estimation of Articulated Body Pose

A fundamental task of vision systems is to infer the state of the world given some form of visual observations. From a computational perspective, this often involves facing an ill-posed problem; e.g., information is lost via projection of the 3D world into a 2D image. Solution of an ill-posed problem requires additional information, usually provided as a model of the underlying process. It is important that the model be both computationally feasible as well as theoretically well-founded. In this thesis, a probabilistic, nonlinear supervised computational learning model is proposed: the Specialized Mappings Architecture (SMA). The SMA framework is demonstrated in a computer vision system that can estimate the articulated pose parameters of a human body or human hands, given images obtained via one or more uncalibrated cameras. The SMA consists of several specialized forward mapping functions that are estimated automatically from training data, and a possibly known feedback function. Each specialized function maps certain domains of the input space (e.g., image features) onto the output space (e.g., articulated body parameters). A probabilistic model for the architecture is first formalized. Solutions to key algorithmic problems are then derived: simultaneous learning of the specialized domains along with the mapping functions, as well as performing inference given inputs and a feedback function. The SMA employs a variant of the Expectation-Maximization algorithm and approximate inference. The approach allows the use of alternative conditional independence assumptions for learning and inference, which are derived from a forward model and a feedback model. Experimental validation of the proposed approach is conducted in the task of estimating articulated body pose from image silhouettes. Accuracy and stability of the SMA framework is tested using artificial data sets, as well as synthetic and real video sequences of human bodies and hands.

Assessing the Security of a Clean-Slate Internet Architecture

The TCP/IP architecture was originally designed without taking security measures into consideration. Over the years, it has been subjected to many attacks, which has led to many patches to counter them. Our investigations into the fundamental principles of networking have shown that carefully following an abstract model of Interprocess Communication (IPC) addresses many problems [1]. Guided by this IPC principle, we designed a clean-slate Recursive INternet Architecture (RINA) [2]. In this paper, we show how, without the aid of cryptographic techniques, the bare-bones architecture of RINA can resist most of the security attacks faced by TCP/IP. We also show how hard it is for an intruder to compromise RINA. Then, we show how RINA inherently supports security policies in a more manageable, on-demand basis, in contrast to the rigid, piecemeal approach of TCP/IP.

Estimating Human Body Pose from a Single Image via the Specialized Mappings Architecture

A non-linear supervised learning architecture, the Specialized Mapping Architecture (SMA) and its application to articulated body pose reconstruction from single monocular images is described. The architecture is formed by a number of specialized mapping functions, each of them with the purpose of mapping certain portions (connected or not) of the input space, and a feedback matching process. A probabilistic model for the architecture is described along with a mechanism for learning its parameters. The learning problem is approached using a maximum likelihood estimation framework; we present Expectation Maximization (EM) algorithms for two different instances of the likelihood probability. Performance is characterized by estimating human body postures from low level visual features, showing promising results.

Biased ART: A Neural Architecture that Shifts Attention Toward Previously Disregarded Features Following an Incorrect Prediction

Memories in Adaptive Resonance Theory (ART) networks are based on matched patterns that focus attention on those portions of bottom-up inputs that match active top-down expectations. While this learning strategy has proved successful for both brain models and applications, computational examples show that attention to early critical features may later distort memory representations during online fast learning. For supervised learning, biased ARTMAP (bARTMAP) solves the problem of over-emphasis on early critical features by directing attention away from previously attended features after the system makes a predictive error. Small-scale, hand-computed analog and binary examples illustrate key model dynamics. Twodimensional simulation examples demonstrate the evolution of bARTMAP memories as they are learned online. Benchmark simulations show that featural biasing also improves performance on large-scale examples. One example, which predicts movie genres and is based, in part, on the Netflix Prize database, was developed for this project. Both first principles and consistent performance improvements on all simulation studies suggest that featural biasing should be incorporated by default in all ARTMAP systems. Benchmark datasets and bARTMAP code are available from the CNS Technology Lab Website: http://techlab.bu.edu/bART/.

Orientational Harmonic Model for Illusory Boundary Formation in Biological Vision

An extension to the Boundary Contour System model is proposed to account for boundary completion through vertices with arbitrary numbers of orientations, in a manner consistent with psychophysical observartions, by way of harmonic resonance in a neural architecture.

A Neural Network Architecture for Figure-ground Separation of Connected Scenic Figures

A neural network model, called an FBF network, is proposed for automatic parallel separation of multiple image figures from each other and their backgrounds in noisy grayscale or multi-colored images. The figures can then be processed in parallel by an array of self-organizing Adaptive Resonance Theory (ART) neural networks for automatic target recognition. An FBF network can automatically separate the disconnected but interleaved spirals that Minsky and Papert introduced in their book Perceptrons. The network's design also clarifies why humans cannot rapidly separate interleaved spirals, yet can rapidly detect conjunctions of disparity and color, or of disparity and motion, that distinguish target figures from surrounding distractors. Figure-ground separation is accomplished by iterating operations of a Feature Contour System (FCS) and a Boundary Contour System (BCS) in the order FCS-BCS-FCS, hence the term FBF, that have been derived from an analysis of biological vision. The FCS operations include the use of nonlinear shunting networks to compensate for variable illumination and nonlinear diffusion networks to control filling-in. A key new feature of an FBF network is the use of filling-in for figure-ground separation. The BCS operations include oriented filters joined to competitive and cooperative interactions designed to detect, regularize, and complete boundaries in up to 50 percent noise, while suppressing the noise. A modified CORT-X filter is described which uses both on-cells and off-cells to generate a boundary segmentation from a noisy image.

A Unified Model of Spatiotemporal Processing in the Retina

A computational model of visual processing in the vertebrate retina provides a unified explanation of a range of data previously treated by disparate models. Three results are reported here: the model proposes a functional explanation for the primary feed-forward retinal circuit found in vertebrate retinae, it shows how this retinal circuit combines nonlinear adaptation with the desirable properties of linear processing, and it accounts for the origin of parallel transient (nonlinear) and sustained (linear) visual processing streams as simple variants of the same retinal circuit. The retina, owing to its accessibility and to its fundamental role in the initial transduction of light into neural signals, is among the most extensively studied neural structures in the nervous system. Since the pioneering anatomical work by Ramón y Cajal at the turn of the last century[1], technological advances have abetted detailed descriptions of the physiological, pharmacological, and functional properties of many types of retinal cells. However, the relationship between structure and function in the retina is still poorly understood. This article outlines a computational model developed to address fundamental constraints of biological visual systems. Neurons that process nonnegative input signals-such as retinal illuminance-are subject to an inescapable tradeoff between accurate processing in the spatial and temporal domains. Accurate processing in both domains can be achieved with a model that combines nonlinear mechanisms for temporal and spatial adaptation within three layers of feed-forward processing. The resulting architecture is structurally similar to the feed-forward retinal circuit connecting photoreceptors to retinal ganglion cells through bipolar cells. This similarity suggests that the three-layer structure observed in all vertebrate retinae[2] is a required minimal anatomy for accurate spatiotemporal visual processing. This hypothesis is supported through computer simulations showing that the model's output layer accounts for many properties of retinal ganglion cells[3],[4],[5],[6]. Moreover, the model shows how the retina can extend its dynamic range through nonlinear adaptation while exhibiting seemingly linear behavior in response to a variety of spatiotemporal input stimuli. This property is the basis for the prediction that the same retinal circuit can account for both sustained (X) and transient (Y) cat ganglion cells[7] by simple morphological changes. The ability to generate distinct functional behaviors by simple changes in cell morphology suggests that different functional pathways originating in the retina may have evolved from a unified anatomy designed to cope with the constraints of low-level biological vision.

A Spectral Network Model of Pitch Perception

A model of pitch perception, called the Spatial Pitch Network or SPINET model, is developed and analyzed. The model neurally instantiates ideas front the spectral pitch modeling literature and joins them to basic neural network signal processing designs to simulate a broader range of perceptual pitch data than previous spectral models. The components of the model arc interpreted as peripheral mechanical and neural processing stages, which arc capable of being incorporated into a larger network architecture for separating multiple sound sources in the environment. The core of the new model transforms a spectral representation of an acoustic source into a spatial distribution of pitch strengths. The SPINET model uses a weighted "harmonic sieve" whereby the strength of activation of a given pitch depends upon a weighted sum of narrow regions around the harmonics of the nominal pitch value, and higher harmonics contribute less to a pitch than lower ones. Suitably chosen harmonic weighting functions enable computer simulations of pitch perception data involving mistuned components, shifted harmonics, and various types of continuous spectra including rippled noise. It is shown how the weighting functions produce the dominance region, how they lead to octave shifts of pitch in response to ambiguous stimuli, and how they lead to a pitch region in response to the octave-spaced Shepard tone complexes and Deutsch tritones without the use of attentional mechanisms to limit pitch choices. An on-center off-surround network in the model helps to produce noise suppression, partial masking and edge pitch. Finally, it is shown how peripheral filtering and short term energy measurements produce a model pitch estimate that is sensitive to certain component phase relationships.