Spatiotemporal Gesture Segmentation

Spotting patterns of interest in an input signal is a very useful task in many different fields including medicine, bioinformatics, economics, speech recognition and computer vision. Example instances of this problem include spotting an object of interest in an image (e.g., a tumor), a pattern of interest in a time-varying signal (e.g., audio analysis), or an object of interest moving in a specific way (e.g., a human's body gesture). Traditional spotting methods, which are based on Dynamic Time Warping or hidden Markov models, use some variant of dynamic programming to register the pattern and the input while accounting for temporal variation between them. At the same time, those methods often suffer from several shortcomings: they may give meaningless solutions when input observations are unreliable or ambiguous, they require a high complexity search across the whole input signal, and they may give incorrect solutions if some patterns appear as smaller parts within other patterns. In this thesis, we develop a framework that addresses these three problems, and evaluate the framework's performance in spotting and recognizing hand gestures in video. The first contribution is a spatiotemporal matching algorithm that extends the dynamic programming formulation to accommodate multiple candidate hand detections in every video frame. The algorithm finds the best alignment between the gesture model and the input, and simultaneously locates the best candidate hand detection in every frame. This allows for a gesture to be recognized even when the hand location is highly ambiguous. The second contribution is a pruning method that uses model-specific classifiers to reject dynamic programming hypotheses with a poor match between the input and model. Pruning improves the efficiency of the spatiotemporal matching algorithm, and in some cases may improve the recognition accuracy. The pruning classifiers are learned from training data, and cross-validation is used to reduce the chance of overpruning. The third contribution is a subgesture reasoning process that models the fact that some gesture models can falsely match parts of other, longer gestures. By integrating subgesture reasoning the spotting algorithm can avoid the premature detection of a subgesture when the longer gesture is actually being performed. Subgesture relations between pairs of gestures are automatically learned from training data. The performance of the approach is evaluated on two challenging video datasets: hand-signed digits gestured by users wearing short sleeved shirts, in front of a cluttered background, and American Sign Language (ASL) utterances gestured by ASL native signers. The experiments demonstrate that the proposed method is more accurate and efficient than competing approaches. The proposed approach can be generally applied to alignment or search problems with multiple input observations, that use dynamic programming to find a solution.

Object Detection at the Optimal Scale with Hidden State Shape Models

Hidden State Shape Models (HSSMs) [2], a variant of Hidden Markov Models (HMMs) [9], were proposed to detect shape classes of variable structure in cluttered images. In this paper, we formulate a probabilistic framework for HSSMs which provides two major improvements in comparison to the previous method [2]. First, while the method in [2] required the scale of the object to be passed as an input, the method proposed here estimates the scale of the object automatically. This is achieved by introducing a new term for the observation probability that is based on a object-clutter feature model. Second, a segmental HMM [6, 8] is applied to model the "duration probability" of each HMM state, which is learned from the shape statistics in a training set and helps obtain meaningful registration results. Using a segmental HMM provides a principled way to model dependencies between the scales of different parts of the object. In object localization experiments on a dataset of real hand images, the proposed method significantly outperforms the method of [2], reducing the incorrect localization rate from 40% to 15%. The improvement in accuracy becomes more significant if we consider that the method proposed here is scale-independent, whereas the method of [2] takes as input the scale of the object we want to localize.

A Model of Cerebellar Adaptation of Grip Forces During Lifting

We investigated adaptive neural control of precision grip forces during object lifting. A model is presented that adjusts reactive and anticipatory grip forces to a level just above that needed to stabilize lifted objects in the hand. The model obeys priciples of cerebellar structure and function by using slip sensations as error signals to adapt phasic motor commands to tonic force generators associated with output synergies controlling grip aperture. The learned phasic commands are weight and texture-dependent. Simulations of the new curcuit model reproduce key aspects of experimental observations of force application. Over learning trials, the onset of grip force buildup comes to lead the load force buildup, and the rate-of-rise of grip force, but not load force, scales inversely with the friction of the gripped object.

A Quantitative Evaluation of the AVITEWRITE Model of Handwriting Learning

Much sensory-motor behavior develops through imitation, as during the learning of handwriting by children. Such complex sequential acts are broken down into distinct motor control synergies, or muscle groups, whose activities overlap in time to generate continuous, curved movements that obey an intense relation between curvature and speed. The Adaptive Vector Integration to Endpoint (AVITEWRITE) model of Grossberg and Paine (2000) proposed how such complex movements may be learned through attentive imitation. The model suggest how frontal, parietal, and motor cortical mechanisms, such as difference vector encoding, under volitional control from the basal ganglia, interact with adaptively-timed, predictive cerebellar learning during movement imitation and predictive performance. Key psycophysical and neural data about learning to make curved movements were simulated, including a decrease in writing time as learning progresses; generation of unimodal, bell-shaped velocity profiles for each movement synergy; size scaling with isochrony, and speed scaling with preservation of the letter shape and the shapes of the velocity profiles; an inverse relation between curvature and tangential velocity; and a Two-Thirds Power Law relation between angular velocity and curvature. However, the model learned from letter trajectories of only one subject, and only qualitative kinematic comparisons were made with previously published human data. The present work describes a quantitative test of AVITEWRITE through direct comparison of a corpus of human handwriting data with the model's performance when it learns by tracing human trajectories. The results show that model performance was variable across subjects, with an average correlation between the model and human data of 89+/-10%. The present data from simulations using the AVITEWRITE model highlight some of its strengths while focusing attention on areas, such as novel shape learning in children, where all models of handwriting and learning of other complex sensory-motor skills would benefit from further research.

Biologically Inspired Approaches to Automated Feature Extraction and Target Recognition

Ongoing research at Boston University has produced computational models of biological vision and learning that embody a growing corpus of scientific data and predictions. Vision models perform long-range grouping and figure/ground segmentation, and memory models create attentionally controlled recognition codes that intrinsically cornbine botton-up activation and top-down learned expectations. These two streams of research form the foundation of novel dynamically integrated systems for image understanding. Simulations using multispectral images illustrate road completion across occlusions in a cluttered scene and information fusion from incorrect labels that are simultaneously inconsistent and correct. The CNS Vision and Technology Labs (cns.bu.edulvisionlab and cns.bu.edu/techlab) are further integrating science and technology through analysis, testing, and development of cognitive and neural models for large-scale applications, complemented by software specification and code distribution.

SOVEREIGN: A Self-Organizing, Vision, Expectation, Recognition, Emotion, Intelligent, Goal-Oriented Navigation System

Both animals and mobile robots, or animats, need adaptive control systems to guide their movements through a novel environment. Such control systems need reactive mechanisms for exploration, and learned plans to efficiently reach goal objects once the environment is familiar. How reactive and planned behaviors interact together in real time, and arc released at the appropriate times, during autonomous navigation remains a major unsolved problern. This work presents an end-to-end model to address this problem, named SOVEREIGN: A Self-Organizing, Vision, Expectation, Recognition, Emotion, Intelligent, Goal-oriented Navigation system. The model comprises several interacting subsystems, governed by systems of nonlinear differential equations. As the animat explores the environment, a vision module processes visual inputs using networks that arc sensitive to visual form and motion. Targets processed within the visual form system arc categorized by real-time incremental learning. Simultaneously, visual target position is computed with respect to the animat's body. Estimates of target position activate a motor system to initiate approach movements toward the target. Motion cues from animat locomotion can elicit orienting head or camera movements to bring a never target into view. Approach and orienting movements arc alternately performed during animat navigation. Cumulative estimates of each movement, based on both visual and proprioceptive cues, arc stored within a motor working memory. Sensory cues are stored in a parallel sensory working memory. These working memories trigger learning of sensory and motor sequence chunks, which together control planned movements. Effective chunk combinations arc selectively enhanced via reinforcement learning when the animat is rewarded. The planning chunks effect a gradual transition from reactive to planned behavior. The model can read-out different motor sequences under different motivational states and learns more efficient paths to rewarded goals as exploration proceeds. Several volitional signals automatically gate the interactions between model subsystems at appropriate times. A 3-D visual simulation environment reproduces the animat's sensory experiences as it moves through a simplified spatial environment. The SOVEREIGN model exhibits robust goal-oriented learning of sequential motor behaviors. Its biomimctic structure explicates a number of brain processes which are involved in spatial navigation.

SOVEREIGN: An Autonomous Neural System for Incrementally Learning Planned Action Sequence to Navigate Towards a Rewarded Goal

How do reactive and planned behaviors interact in real time? How are sequences of such behaviors released at appropriate times during autonomous navigation to realize valued goals? Controllers for both animals and mobile robots, or animats, need reactive mechanisms for exploration, and learned plans to reach goal objects once an environment becomes familiar. The SOVEREIGN (Self-Organizing, Vision, Expectation, Recognition, Emotion, Intelligent, Goaloriented Navigation) animat model embodies these capabilities, and is tested in a 3D virtual reality environment. SOVEREIGN includes several interacting subsystems which model complementary properties of cortical What and Where processing streams and which clarify similarities between mechanisms for navigation and arm movement control. As the animat explores an environment, visual inputs are processed by networks that are sensitive to visual form and motion in the What and Where streams, respectively. Position-invariant and sizeinvariant recognition categories are learned by real-time incremental learning in the What stream. Estimates of target position relative to the animat are computed in the Where stream, and can activate approach movements toward the target. Motion cues from animat locomotion can elicit head-orienting movements to bring a new target into view. Approach and orienting movements are alternately performed during animat navigation. Cumulative estimates of each movement are derived from interacting proprioceptive and visual cues. Movement sequences are stored within a motor working memory. Sequences of visual categories are stored in a sensory working memory. These working memories trigger learning of sensory and motor sequence categories, or plans, which together control planned movements. Predictively effective chunk combinations are selectively enhanced via reinforcement learning when the animat is rewarded. Selected planning chunks effect a gradual transition from variable reactive exploratory movements to efficient goal-oriented planned movement sequences. Volitional signals gate interactions between model subsystems and the release of overt behaviors. The model can control different motor sequences under different motivational states and learns more efficient sequences to rewarded goals as exploration proceeds.

Spikes, Synchrony, and Attentive Learning by Laminar Thalamocortical Circuits

This article develops the Synchronous Matching Adaptive Resonance Theory (SMART) neural model to explain how the brain may coordinate multiple levels of thalamocortical and corticocortical processing to rapidly learn, and stably remember, important information about a changing world. The model clarifies how bottom-up and top-down processes work together to realize this goal, notably how processes of learning, expectation, attention, resonance, and synchrony are coordinated. The model hereby clarifies, for the first time, how the following levels of brain organization coexist to realize cognitive processing properties that regulate fast learning and stable memory of brain representations: single cell properties, such as spiking dynamics, spike-timing-dependent plasticity (STDP), and acetylcholine modulation; detailed laminar thalamic and cortical circuit designs and their interactions; aggregate cell recordings, such as current-source densities and local field potentials; and single cell and large-scale inter-areal oscillations in the gamma and beta frequency domains. In particular, the model predicts how laminar circuits of multiple cortical areas interact with primary and higher-order specific thalamic nuclei and nonspecific thalamic nuclei to carry out attentive visual learning and information processing. The model simulates how synchronization of neuronal spiking occurs within and across brain regions, and triggers STDP. Matches between bottom-up adaptively filtered input patterns and learned top-down expectations cause gamma oscillations that support attention, resonance, and learning. Mismatches inhibit learning while causing beta oscillations during reset and hypothesis testing operations that are initiated in the deeper cortical layers. The generality of learned recognition codes is controlled by a vigilance process mediated by acetylcholine.

Cortical Learning of Recognition Categories: A Resolution of the Exemplar Vs. Prototype Debate

Do humans and animals learn exemplars or prototypes when they categorize objects and events in the world? How are different degrees of abstraction realized through learning by neurons in inferotemporal and prefrontal cortex? How do top-down expectations influence the course of learning? Thirty related human cognitive experiments (the 5-4 category structure) have been used to test competing views in the prototype-exemplar debate. In these experiments, during the test phase, subjects unlearn in a characteristic way items that they had learned to categorize perfectly in the training phase. Many cognitive models do not describe how an individual learns or forgets such categories through time. Adaptive Resonance Theory (ART) neural models provide such a description, and also clarify both psychological and neurobiological data. Matching of bottom-up signals with learned top-down expectations plays a key role in ART model learning. Here, an ART model is used to learn incrementally in response to 5-4 category structure stimuli. Simulation results agree with experimental data, achieving perfect categorization in training and a good match to the pattern of errors exhibited by human subjects in the testing phase. These results show how the model learns both prototypes and certain exemplars in the training phase. ART prototypes are, however, unlike the ones posited in the traditional prototype-exemplar debate. Rather, they are critical patterns of features to which a subject learns to pay attention based on past predictive success and the order in which exemplars are experienced. Perturbations of old memories by newly arriving test items generate a performance curve that closely matches the performance pattern of human subjects. The model also clarifies exemplar-based accounts of data concerning amnesia.

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/.

Task-Irrelevant Perceptual Learning Specific to the Contrast Polarity of Motion Stimuli

Studies of perceptual learning have focused on aspects of learning that are related to early stages of sensory processing. However, conclusions that perceptual learning results in low-level sensory plasticity are of great controversy, largely because such learning can often be attributed to plasticity in later stages of sensory processing or in the decision processes. To address this controversy, we developed a novel random dot motion (RDM) stimulus to target motion cells selective to contrast polarity, by ensuring the motion direction information arises only from signal dot onsets and not their offsets, and used these stimuli in conjunction with the paradigm of task-irrelevant perceptual learning (TIPL). In TIPL, learning is achieved in response to a stimulus by subliminally pairing that stimulus with the targets of an unrelated training task. In this manner, we are able to probe learning for an aspect of motion processing thought to be a function of directional V1 simple cells with a learning procedure that dissociates the learned stimulus from the decision processes relevant to the training task. Our results show learning for the exposed contrast polarity and that this learning does not transfer to the unexposed contrast polarity. These results suggest that TIPL for motion stimuli may occur at the stage of directional V1 simple cells.

Cortical Dynamics of Contextually-Cued Attentive Visual Learning and Search: Spatial and Object Evidence Accumulation

How do humans use predictive contextual information to facilitate visual search? How are consistently paired scenic objects and positions learned and used to more efficiently guide search in familiar scenes? For example, a certain combination of objects can define a context for a kitchen and trigger a more efficient search for a typical object, such as a sink, in that context. A neural model, ARTSCENE Search, is developed to illustrate the neural mechanisms of such memory-based contextual learning and guidance, and to explain challenging behavioral data on positive/negative, spatial/object, and local/distant global cueing effects during visual search. The model proposes how global scene layout at a first glance rapidly forms a hypothesis about the target location. This hypothesis is then incrementally refined by enhancing target-like objects in space as a scene is scanned with saccadic eye movements. The model clarifies the functional roles of neuroanatomical, neurophysiological, and neuroimaging data in visual search for a desired goal object. In particular, the model simulates the interactive dynamics of spatial and object contextual cueing in the cortical What and Where streams starting from early visual areas through medial temporal lobe to prefrontal cortex. After learning, model dorsolateral prefrontal cortical cells (area 46) prime possible target locations in posterior parietal cortex based on goalmodulated percepts of spatial scene gist represented in parahippocampal cortex, whereas model ventral prefrontal cortical cells (area 47/12) prime possible target object representations in inferior temporal cortex based on the history of viewed objects represented in perirhinal cortex. The model hereby predicts how the cortical What and Where streams cooperate during scene perception, learning, and memory to accumulate evidence over time to drive efficient visual search of familiar scenes.

Grid Cell Hexagonal Patterns Formed by Fast Self-Organized Learning within Entorhinal Cortex

Grid cells in the dorsal segment of the medial entorhinal cortex (dMEC) show remarkable hexagonal activity patterns, at multiple spatial scales, during spatial navigation. How these hexagonal patterns arise has excited intense interest. It has previously been shown how a selforganizing map can convert firing patterns across entorhinal grid cells into hippocampal place cells that are capable of representing much larger spatial scales. Can grid cell firing fields also arise during navigation through learning within a self-organizing map? A neural model is proposed that converts path integration signals into hexagonal grid cell patterns of multiple scales. This GRID model creates only grid cell patterns with the observed hexagonal structure, predicts how these hexagonal patterns can be learned from experience, and can process biologically plausible neural input and output signals during navigation. These results support a unified computational framework for explaining how entorhinal-hippocampal interactions support spatial navigation.

Fusion Artmap: A Neural Network Architecture for Multi-Channel Data Fusion and Classification

Fusion ARTMAP is a self-organizing neural network architecture for multi-channel, or multi-sensor, data fusion. Single-channel Fusion ARTMAP is functionally equivalent to Fuzzy ART during unsupervised learning and to Fuzzy ARTMAP during supervised learning. The network has a symmetric organization such that each channel can be dynamically configured to serve as either a data input or a teaching input to the system. An ART module forms a compressed recognition code within each channel. These codes, in turn, become inputs to a single ART system that organizes the global recognition code. When a predictive error occurs, a process called paraellel match tracking simultaneously raises vigilances in multiple ART modules until reset is triggered in one of them. Parallel match tracking hereby resets only that portion of the recognition code with the poorest match, or minimum predictive confidence. This internally controlled selective reset process is a type of credit assignment that creates a parsimoniously connected learned network. Fusion ARTMAP's multi-channel coding is illustrated by simulations of the Quadruped Mammal database.

Rule Extraction, Fuzzy ARTMAP, and Medical Databases

This paper shows how knowledge, in the form of fuzzy rules, can be derived from a self-organizing supervised learning neural network called fuzzy ARTMAP. Rule extraction proceeds in two stages: pruning removes those recognition nodes whose confidence index falls below a selected threshold; and quantization of continuous learned weights allows the final system state to be translated into a usable set of rules. Simulations on a medical prediction problem, the Pima Indian Diabetes (PID) database, illustrate the method. In the simulations, pruned networks about 1/3 the size of the original actually show improved performance. Quantization yields comprehensible rules with only slight degradation in test set prediction performance.