Search

In this paper, we apply natural language processing methods to develop models for predicting physiological effects of chemical substances based on their molecular structures. Using string representations of structure as a starting point, we vectorize molecules using two different approaches resulting in sparse and dense vector representations, respectively. We use these representations to train predictive models for a variety of physiological effects such as toxicity, cell cycle arrest and proliferation. Using standard chemical datasets, we empirically demonstrate that such models can achieve high predictive accu-racy.

Agent-based modeling is a paradigm of modeling dynamic systems of interacting agents that are individually governed by specified behavioral rules. Training a model of such agents to produce an emergent behavior by specification of the emergent (as opposed to agent) behavior is easier from a demonstration perspective. While many approaches involve manual behavior specification via code or reliance on a defined taxonomy of possible behaviors, the meta-modeling framework in Miner [2010] generates mapping functions between agent-level parameters and swarm-level parameters, which are re-usable once generated. This work builds on that framework by integrating demonstration by image or video. The demonstrator specifies spatial motion of the agents over time and retrieves agent-level parameters required to execute that motion. The framework, at its core, uses computationally cheap image-processing algorithms. Our work is tested with a combination of primitive visual feature extraction methods (contour area and shape) and features generated using a pre-trained deep neural network in different stages of image featurization. The framework is also evaluated for its potential using complex visual features for all image featurization stages. Experimental results show significant coherence between demonstrated behavior and predicted behavior based on estimated agent-level parameters specific to the spatial arrangement of agents.

In autonomous embedded systems, it is often vital to reduce the amount of actions taken in the real world and energy required to learn a policy. Training reinforcement learning agents from high dimensional image representations can be very expensive and time consuming. Autoencoders are deep neural network used to compress high dimensional data such as pixelated images into small latent representations. This compression model is vital to efficiently learn policies, especially when learning on embedded systems. We have implemented this model on the NVIDIA Jetson TX2 embedded GPU, and evaluated the power consumption, throughput, and energy consumption of the autoencoders for various CPU/GPU core combinations, frequencies, and model parameters. Additionally, we have shown the reconstructions generated by the autoencoder to analyze the quality of the generated compressed representation and also the performance of the reinforcement learning agent. Finally, we have presented an assessment of the viability of training these models on embedded systems and their usefulness in developing autonomous policies. Using autoencoders, we were able to achieve 4-5 $\times$ improved performance compared to a baseline RL agent with a convolutional feature extractor, while using less than 2W of power.

27 International Conference on Case-Based Reasoning, 2019, Otzenhausen, Germany, Deep Learning methods have shown a rapid increase in popularity due to their state-of-the-art performance on many machine learning tasks. However, these methods often rely on extremely large datasets to accurately train the underlying machine learning models. For supervised learning techniques, the human effort required to acquire, encode, and label a sufficiently large dataset may add such a high cost that deploying the algorithms is infeasible. Even if a sufficient workforce exists to create such a dataset, the human annotators may differ in the quality, consistency, and level of granularity of their labels. Any impact this has on the overall dataset quality will ultimately impact the potential performance of an algorithm trained on it. This paper partially addresses this issue by providing an approach, called NOD-CC, for discovering novel object types in images using a combination of Convolutional Neural Networks (CNNs) and Case-Based Reasoning (CBR). The CNN component labels instances of known object types while deferring to the CBR component to identify and label novel, or poorly understood, object types. Thus, our approach leverages the state-of-the-art performance of CNNs in situations where sufficient high-quality training data exists, while minimizing its limitations in data-poor situations. We empirically evaluate our approach on a popular computer vision dataset and show significant improvements to objects classification performance when full knowledge of potential class labels is not known in advance.

Agent-based modeling is a paradigm of modeling dynamic systems of interacting agents that are individually governed by specified behavioral rules. Training a model of such agents to produce an emergent behavior by specification of the emergent (as opposed to agent) behavior is easier from a demonstration perspective. Without the involvement of manual behavior specification via code or reliance on a defined taxonomy of possible behaviors, the demonstrator specifies spatial motion of the agents over time, and retrieves agent-level parameters required to execute that motion. A framework for reproducing emergent behavior, given an abstract demonstration, is discussed in existing work. Our work extends that framework by refining the data that is aggregated to produce the agent-level parameters that the framework provides to the demonstrator. This is done using pruning and outlier detection based on information that is intrinsic to those data points (their source). Using pruning and outlier detection shows potential to refine the aggregation data to a fraction of its size, while maintaining or potentially improving performance in replication of demonstrations.

Limb repositioning is necessary for individuals with severe physical disabilities to sustain muscle strength and prevent pressure sores. As robotic technologies become ubiquitous, these tools offer promise to support the repositioning process. However, research has yet to focus on ways in which individuals with severe physical disabilities can control robots for these tasks. This paper presents a study that examines the needs and attitudes of potential users with physical disabilities to control a robotic aid for limb repositioning. Subjects expressed interest in using brain---computer interface (BCI) and speech recognition technologies for purposes of executing robotic tasks. The performance of four subjects controlling arm movements on an avatar through the keyboard, mouse, BCI, and Dragon NaturallySpeaking speech recognition was evaluated. Although BCI and speech technologies may limit physical fatigue, more challenges were faced using BCI and speech conditions compared to the keyboard and mouse. This research promotes accessibility into mainstream robotic technologies and represents the first step in the development of a robotic prototype using a BCI and speech recognition technologies for limb repositioning.

Denoising data is a preprocessing step for several time series mining algorithms. This step is especially important if the noise in data originates from diverse sources. Consequently, it is commonly used in biomedical applications that use Electroencephalography (EEG) data. In EEG data noise can occur due to ocular, muscular and cardiac activities. In this paper, we explicitly learn to remove noise from time series data without assuming a prior distribution of noise. We propose an online, fully automated, end-to-end system for denoising time series data. Our model for denoising time series is trained using unpaired training corpora and does not need information about the source of the noise or how it is manifested in the time series. We propose a new architecture called AsymmetricGAN that uses a generative adversarial network for denoising time series data. To analyze our approach, we create a synthetic dataset that is easy to visualize and interpret. We also evaluate and show the effectiveness of our approach on an existing EEG dataset.

This brief presents a low-power, flexible, and multichannel electroencephalography (EEG) feature extractor and classifier for the purpose of personalized seizure detection. Various features and classifiers were explored with the goal of maximizing detection accuracy while minimizing power, area, and latency. Additionally, algorithmic and hardware optimizations were identified to further improve performance. The classifiers studied include $k$ -nearest neighbor, support vector machines, naive Bayes, and logistic regression (LR) . All feature and classifier pairs were able to obtain F1 scores over 80% and onset sensitivity of 100% when tested on ten patients. A fully flexible hardware system was implemented that offers parameters for the number of EEG channels, the number of features, the classifier type, and various word width resolutions. Five seizure detection processors with different classifiers have been fully placed and routed on a Virtex-5 field-programmable gate array and been compared. It was found that five features per channel with LR proved to be the best solution for the application of personalized seizure detection. LR had the best average F1 score of 91%, the smallest area and power footprint, and the lowest latency. The ASIC implementation of the same combination in 65-nm CMOS shows that the processor occupies 0.008 mm2 and dissipates 19 nJ at 484 Hz.

Neurophysiological signals such as Electroencephalogram (EEG) can be used in a variety of purposes including detecting fatigue, stress, brain disorders, brain-computer interfaces (BCIs), or building better models of human variability and human brain. However, EEG signal is frequently defiled with other electrical sources not related to brain activity. These artifacts may emerge from external sources such as eye blinks, muscle & head movement, and power lines. Existing techniques for removing artifacts such as independent component analysis (ICA) are highly computationally intensive and need to process data offline due in large part to the highly volatile, non-stationary nature of the brain signals, especially in natural, noise-ridden environments.

Traditional approaches to building a large scale knowledge graph have usually relied on extracting information (entities, their properties, and relations between them) from unstructured text (e.g. Dbpedia). Recent advances in Convolutional Neural Networks (CNN) allow us to shift our focus to learning entities and relations from images, as they build robust models that require little or no pre-processing of the images. In this paper, we present an approach to identify and extract spatial relations (e.g., The girl is standing behind the table) from images using CNNs. Our research addresses two specific challenges: providing insight into how spatial relations are learned by the network and which parts of the image are used to predict these relations. We use the pre-trained network VGGNet to extract features from an image and train a Multi-layer Perceptron (MLP) on a set of synthetic images and the sun09 dataset to extract spatial relations. The MLP predicts spatial relations without a bounding box around the objects or the space in the image depicting the relation. To understand how the spatial relations are represented in the network, a heatmap is overlayed on the image to show the regions that are deemed important by the network. Also, we analyze the MLP to show the relationship between the activation of consistent groups of nodes and the prediction of a spatial relation. We show how the loss of these groups affects the network's ability to identify relations.

We propose a simple but strong baseline for time series classification from scratch with deep neural networks. Our proposed baseline models are pure end-to-end without any heavy preprocessing on the raw data or feature crafting. The proposed Fully Convolutional Network (FCN) achieves premium performance to other state-of-the-art approaches and our exploration of the very deep neural networks with the ResNet structure is also competitive. The global average pooling in our convolutional model enables the exploitation of the Class Activation Map (CAM) to find out the contributing region in the raw data for the specific labels. Our models provides a simple choice for the real world application and a good starting point for the future research. An overall analysis is provided to discuss the generalization capability of our models, learned features, network structures and the classification semantics.

The increase in available spatial trajectory data has led to a massive amount of geo-positioned data that can be exploited to improve understanding of human behavior. However, the noisy nature and massive size of the data make it difficult to extract meaningful trajectory features. In this work, a context-free grammar representation of spatial trajectories is employed to discover frequent segments or motifs within trajectories. Additionally, a set of basis motifs is developed that defines all movement characteristics among a set of trajectories, which can be used to evaluate patterns within a trajectory (intra-trajectory) and between multiple trajectories (inter-trajectory). The approach is realized and demonstrable through the Symbolic Trajectory Analysis and VIsualization System (STAVIS) 2.0, which performs grammar inference on spatial trajectories, mines motifs, and discovers various pattern sets through motif-based analysis.

Agent-based modeling is a paradigm of modeling dynamic systems of interacting agents that are individually governed by specified behavioral rules. Training a model of such agents to produce an emergent behavior by specification of the emergent (as opposed to agent) behavior is easier from a demonstration perspective. While many approaches involve manual behavior specification via code or reliance on a defined taxonomy of possible behaviors, the meta-modeling framework in existing work generates mapping functions between agent-level parameters and swarm-level parameters which are re-usable once generated. This work builds on that framework by integrating demonstration by image or video. The demonstrator specifies spatial motion of the agents over time, and retrieves agent-level parameters required to execute that motion. The framework, at its core, uses computationally cheap image processing algorithms. Our work is tested with a combination of primitive visual feature extraction methods (contour area and shape) and features generated using a pre-trained deep neural network in different stages of image featurization. Experimental results show significant coherence between demonstrated behavior and predicted behavior based on estimated agent-level parameters specific to a class of behaviors (spatial arrangement of agents).

The problem of Learning from Demonstration is targeted at learning to perform tasks based on observed examples. One approach to Learning from Demonstration is Inverse Reinforcement Learning, in which actions are observed to infer rewards. This work combines a feature based state evaluation approach to Inverse Reinforcement Learning with neuroevolution, a paradigm for modifying neural networks based on their performance on a given task. Neural networks are used to learn from a demonstrated expert policy and are evolved to generate a policy similar to the demonstration. The algorithm is discussed and evaluated against competitive feature-based Inverse Reinforcement Learning approaches. At the cost of execution time, neural networks allow for non-linear combinations of features in state evaluations. These valuations may correspond to state value or state reward. This results in better correspondence to observed examples as opposed to using linear combinations. This work also extends existing work on Bayesian Non-Parametric Feature Construction for Inverse Reinforcement Learning by using non-linear combinations of intermediate data to improve performance. The algorithm is observed to be specifically suitable for a linearly solvable non-deterministic Markov Decision Processes in which multiple rewards are sparsely scattered in state space. A conclusive performance hierarchy between evaluated algorithms is presented., 12 pages, 15 figures

This work introduces adversarial feature selection, a game between a feature selection agent and its adversary. The adversarial approach is drawn from existing work on adversarial classification. The feature selection algorithm selects a subset of features from the original set based on their utility towards classification accuracy. A cost is incurred based on features selected. The adversary modifies features with less utility as ones with higher utility. In this way, the adversary profits if the feature selector selects these features at increased costs. A base feature selection algorithm is used to generate cost values for features. The problem is formalized and a corresponding model of the algorithm is discussed and evaluated.

Discretization is a crucial first step in several time series mining applications. Our research proposes a novel method to discretize time series data and develops a similarity score based on the discretized representation. The similarity score allows us to compare two time series sequences and enables us to perform pattern learning tasks such as clustering, classification, and anomaly detection. We propose a generative model for discretization based on multiple normal distributions and create an optimization technique to learn parameters of these normal distributions. To show the effectiveness of our approach, we perform comprehensive experiments in classifying datasets from the UCR time series repository.

Maintaining adequate performance in dynamic and uncertain settings has been a perennial stumbling block for intelligent systems. Nevertheless, any system intended for real-world deployment must be able to accommodate unexpected change—that is, it must be perturbation tolerant. We have found that metacognitive monitoring and control—the ability of a system to self-monitor its own decision-making processes and ongoing performance, and to make targeted changes to its beliefs and action-determining components—can play an important role in helping intelligent systems cope with the perturbations that are the inevitable result of real-world deployment. In this article we present the results of several experiments demonstrating the efficacy of metacognition in improving the perturbation tolerance of reinforcement learners, and discuss a general theory of metacognitive monitoring and control, in a form we call the metacognitive loop. ||This research is supported in part by the AFOSR and ONR.

As a field, Grammatical Inference addresses both theoretical and empirical learning problems, and the collection of papers within this special issue attests both to the diversity of these problems as well as the advances and insights that are being made by the researchers working within it. Thus we hope this special issue is of interest to the readership of Machine Learning.

With the advent of reliable and continuously operating noise monitoring systems, we are now faced with an unprecedented amount of noise monitor data. In the context of environmental noise monitoring, there is a need to automatically detect, separate, and classify all environmental noise sources. This is a complex task because sources can overlap, vary by location, and have an unbounded number of noise sources that a monitor device may record. In this study, we synthetically generate datasets that contain Gaussian noise and overlaps for several pre-labeled environmental noise monitoring datasets to examine how well deep learning methods (e.g., autoencoders) can separate environmental noise sources. In addition to examining performance, we also focus on understanding which signal features and separation metrics are useful to this problem.

Signatures are the single most widely used method of identifying an individual but they carry with them an alarmingly significant number of vulnerabilities, implying the need for an effective and robust method of precisely identifying an individual's signature. The signature of an individual is visually acquired by using a pen-based tracking system [1], [2]. This paper considers the possibility of discretizing visually acquired signatures to represent them as an unordered collection of words and then use Sequential Mining Optimization (SMO) for training Support Vector Machines (SVM) to classify the signature as either legitimate or forged. Signature discretization is done using Symbolic Aggregate Approximation (SAX) [4]. SAX reduces the dimensions of the signature and produces a list of SAX words that are then represented as a bag-of-patterns model for classification purposes [6]. The approach was tested on a dataset of 3960 signatures of 106 subjects distributed across two sets. The results show good classification accuracy bolstering the real time application of the methodology.

Robotic technologies can provide people with disabilities the invaluable tools to perform Activities of Daily Living (ADL). Few studies have investigated how effective and accessible the control of robotic aids is for people with severe physical disabilities with respect to their needs and current facility with technology. Though present-day robotic aids can help people with disabilities with important daily living tasks, there is still room for improvement. In this chapter, an overview of the state-of-the-art robotics that help promote independence for people with disabilities, especially individuals with physical limitations, is presented. Active research areas supporting tasks and services for persons with physical disabilities include development of robotic prototypes, designed specifically for domestic applications, rehabilitation, mobility assistance, personal hygiene care and meal assistance, and education and employment. Existing challenges and recommendations to support and increase independence for persons with motor limitations with robotics technologies are provided as insight to advance research.

The problem of frequent and anomalous patterns discovery in time series has received a lot of attention in the past decade. Addressing the common limitation of existing techniques, which require a pattern length to be known in advance, we recently proposed grammar-based algorithms for efficient discovery of variable length frequent and rare patterns. In this paper we present GrammarViz 2.0, an interactive tool that, based on our previous work, implements algorithms for grammar-driven mining and visualization of variable length time series patterns1.

Constructing an image classification system using strong, local invariant descriptors is both time consuming and tedious, requiring many experimentations and parameter tunings to obtain an adequately performing model. Furthermore training a system in a given domain and then migrating the model to a separate domain will likely yield poor performance. As the recent Boston Marathon attacks demonstrated, large, unstructured image databases from traffic cameras, security systems, law enforcement officials, and citizens can be quickly amassed for authorities to review, however, reviewing each and every image is a expensive undertaking, in terms of both time and human intervention. Inherently, reviewing crime scene images is a classification task. For example, authorities may want to know if a given image contains a suspect, a suspicious package, or if there are injured people in the photo. Given an emergency situation, these classifications will be needed as quickly and accurately as possible. In this work we present a rapidly deployable image classification system using "feature-views", which each view consists of a set of weak, global features. These weak global descriptors are computationally simple to extract, intuitive to understand, and require substantially less parameter tuning than their local invariant counterparts. We demonstrate that by combining weak features with ensemble methods we are able to outperform current state-of-the-art methods or achieve comparable accuracy with much less effort and domain knowledge. Finally we provide both theoretical and empirical justification for our ensemble framework that can be used to construct rapidly deployable image classification systems called "Ecosembles".

In this paper, we review recent advances in Reinforcement Learning (RL) in light of potential applications to robotics, introduce the basic concepts of RL and Markov Decision Process (MDP), and compare different RL algorithms such as Q-learning, Temporal Difference learning, the Actor Critic, and the Natural Actor Critic. We conclude that policy gradient methods are more suitable for solving continuous state/action MDP problems than RL with lookup tables or general function approximators. Further, natural policy gradient methods can efficiently converge to locally optimal solutions. Some simulation results are given to support our arguments. We also present a brief overview of our approach to developing an autonomous robot agent that can perceive, learn from and interact with the environment, and reason about and handle unexpected problems using its knowledge base.

Spatial trajectory analysis is crucial to uncovering insights into the motives and nature of human behavior. In this work, we study the problem of discovering motifs in trajectories based on symbolically transformed representations and context free grammars. We propose a fast and robust grammar induction algorithm called mSEQUITUR to infer a grammar rule set from a trajectory for motif generation. Second, we designed the Symbolic Trajectory Analysis and VIsualization System (STAVIS), the first of its kind trajectory analytical system that applies grammar inference to derive trajectory signatures and enable mining tasks on the signatures. Third, an empirical evaluation is performed to demonstrate the efficiency and effectiveness of mSEQUITUR for generating trajectory signatures and discovering motifs.

Deep learning is becoming ubiquitous; it is the underlying and driving force behind many heavily embedded technologies in society (e.g., search engines, fraud detection warning systems, and social-media facial recognition algorithms). Over the past few years there has been a steady increase in the number of audio related applications of deep learning. Recently, Nykaza et al. presented a pedagogical approach to understanding how the hidden layers recreate, separate, and classify environmental noise signals. That work presented some feature extraction examples using simple pure tone, chord, and environmental noise datasets. In this paper, we build upon this recent analysis and expand the datasets to include more realistic representations of those datasets with the inclusion of noise and overlapping signals. Additionally, we consider other related architectures (e.g., variant-autoencoders, recurrent neural networks, and fixing hidden nodes/layers), explore their advantages/drawbacks, and provide insights on ...

Deep learning is becoming ubiquitous; it is the underlying and driving force behind many technologies we use everyday (e.g., search engines, fraud detection warning systems, and social-media facial recognition algorithms). Over the past few years, there has been a steady increase in the number of audio and acoustics related applications of deep learning. But what is exactly going on under the hood? In this paper, we focus on deep learning algorithms for unsupervised feature learning. We take a pedagogical approach to understanding how the hidden layers recreate, separate, and classify audio signals. We begin with a simple pure tone dataset, and systematically increase the complexity of this dataset in both frequency and time. We end the presentation with some feature extraction examples from real-world environmental recordings, and find that these features are easier to interpret given the understanding developed from the simpler tone datasets. The unsupervised feature learning techniques explored in this paper include: restricted Boltzmann machines (RBMs) and auto-encoders (AEs).

More than a decade of research has produced numerous representations and similarity measures to support time series classification and clustering. Yet most of the work in the field is so focused on the representation or similarity measure that it ignores the possibility of improving performance using ensembles of representations or classifiers. This paper explores ways of exploiting representational diversity for time series classification via ensembles of representations. We focus on the Symbolic Aggregate approXimation (SAX) discretization method coupled with the bag-of-patterns (BoP) representation because of their state-of-the-art performance in the single representation/classifier case. Experiments with a number of standard benchmark time series datasets and a new dataset of vital signs collected from patients suffering from traumatic brain injury demonstrate the power of the ensemble approaches. The result is a single method that is often significantly better than vanilla SAX/BoP and compares favorably on a per dataset basis with the best methods reported in the literature for each dataset.

Metacognition is a powerful tool that can play machine learning and commonsense reasoning off one another synergistically. Based on our previous work in metacognition, we present the vision of a long-term project called “Robot in a Room” in this paper. Similar to newborn babies trying to learn about themselves, surroundings and things that they are able to do, this new project involves a mobile robot that does not know about its environment and its own capabilities i.e. it does not start with a robust self-model. However, it has a primary goal: to learn about itself and explore its environment (the room it lives in).

We have initiated a long-term robotics project based on our previous work on metacognition as a powerful tool that can synergistically play machine learning and commonsense reasoning off one another. The new project involves a mobile robot that lives in a room and learns about the room and about itself. The robot is initially set up to have a standard set of facilities (vision, IR, limb, wheels, planners, learning modules, some modest NLP, a reasoner, etc.) but it does not know much about its capabilities or how to properly use them. It has a prime directive: to learn. This paper will focus on one of the first major questions of this project: can we use Growing Neural Gas (GNG) to discover the physical structure of an environment and, if so, what are the limits of its use? To answer this question, we have devised an experiment to test whether our robot can distinguish between two identical objects using only GNG. Preliminary results suggest that passing image data along with robotic control signal data is sufficient to autonomously detect the basic physical structure of a room.

Biometric and soft biometric features can be used to identify people in disaster situations, but the use of biometric features or pictures of victims may lead to privacy issues. Using text-based descriptors to describe disaster victim images would help in making person data public and also in searching for a particular person in a large database. In this paper, work on combining soft biometric features using a Markov network is presented. The proposed structure exploits the relationships between different soft biometric features and results into a robust text descriptor to help in identification of a person from a patient triage image. We show how interaction between individual feature detectors can lead to increased accuracy in the resulting text descriptor.

In this paper we present two novel multivariate time series representations to classify physiological data of different lengths. The representations may be applied to any group of multivariate time series data that examine the state or health of an entity. Multivariate Bag-of-Patterns and Stacked Bags of-Patterns improve on their univariate counterpart, inspired by the bag-of-words model, by using multiple time series and analyzing the data in a multivariate fashion. We also borrow techniques from the natural language processing domain such as term frequency and inverse document frequency to improve classification accuracy. We introduce a technique named inverse frequency and present experimental results on classifying patients who have experienced acute episodes of hypotension.

In this paper we present two novel multivariate time series representations to classify physiological data of different lengths. The representations may be applied to any group of multivariate time series data that examine the state or health of an entity. Multivariate Bag-of-Patterns and Stacked Bags of-Patterns improve on their univariate counterpart, inspired by the bag-of-words model, by using multiple time series and analyzing the data in a multivariate fashion. We also borrow techniques from the natural language processing domain such as term frequency and inverse document frequency to improve classification accuracy. We introduce a technique named inverse frequency and present experimental results on classifying patients who have experienced acute episodes of hypotension.

Text-to-Sketch (T2S) is a class of problems in which geolocation is performed using natural language descriptions of a location or locations as input. This is a challenging problem due to the many sources of uncertainty inherent to the task: there is often syntactic and semantic ambiguity present in the input observations, as well as referential ambiguity when the language used to describe the scene may refer to many possible objects or locations in the world. Tracking problems, in which the Text-to-Sketch paradigm is extended to incorporate multiple locations and movements over a temporal dimension, introduce additional uncertainty. We describe a tool for managing the uncertainty in Text-to-Sketch problems called MUTTS. The MUTTS system combines traditional natural language processing (NLP) tools with algorithms used to manage uncertainty in mobile robot navigation to allow the temporal and geographical constraints in the text to incrementally reduce the overall uncertainty of a subject's location and produce high quality sketches of the subject's location and movements over time.

Computational approaches to language acquisition typically reduce the problem to one of either learning syntax, or learning a lexicon -- isolated tasks. These problem simplifications have led to disconnected solutions and algorithms with no expectation of grounded inputs. We propose a language learning framework for bootstrapping the acquisition of a grammar and learning lexical semantics. We describe the framework and provide an instantiation using context-free grammars and a lambda-calculus semantic representation.

Proceedings of the 4th International Conference on Pervasive Technologies Related to Assistive Environments, With robotics technology, services can be provided to care for individuals with disabilities. This paper describes an eff ort to improve path planning performance for a robotic arm, resulting in faster user response in real-time. For a robotic arm, particularly with multiple degrees of freedom, path planning is computationally expensive. We propose that it is possible to achieve rapid response times with an assistive robotic arm by caching frequent arm trajectories and creating a "roadmap" of arm movements. By calculating trajectories to possible target goals in advance, we anticipate an improvement in user response times.

Time series data naturally arises in many domains, such as industrial process control, robotics, finance, medicine, climatology, and numerous others. In many cases variables known to be causally relevant cannot be measured directly or the existence of such variables is unknown. This paper presents an extension of the neural network architecture, called the LO-net [1], for inferring both the existence and values of hidden variables in streaming multivariate time series, leading to deeper understanding of the domain and more accurate prediction. The core idea is to initially make predictions with one network (the observable or O net) based on a time delay embedding, following this with a gradual reduction in the temporal scope of the embedding that forces a second network (the latent or L net) to learn to approximate the value of a single hidden variable, which is then input to the O net based on the original time delay embedding. Experiments show that the architecture efficiently and accurately identifies the number of hidden variables and their values over time.

Proceedings of the 18th ACM Conference on Information and Knowledge Management, The standard method for combating spam, either in email or on the web, is to train a classifier on manually labeled instances. As the spammers change their tactics, the performance of such classifiers tends to decrease over time. Gathering and labeling more data to periodically retrain the classifier is expensive. We present a method based on an ensemble of classifiers that can detect when its performance might be degrading and retrain itself, all without manual intervention. Experiments with a real-world dataset from the blog domain show that our methods can significantly reduce the number of times classifiers are retrained when compared to a fixed retraining schedule, and they maintain classification accuracy even in the absence of manually labeled examples.

The metacognitive loop (MCL) is architecture for automated noting, repairing, and learning from errors. Initial work on this award involved building special purpose MCL programs for each individual application domain Later in the award period a more uniform approach was designed that employs a general framework providing ontologies for types of Indications, Failures, and Repairs. This allows the past results in different domains to be achievable by a single MCL module, changing only the domain and the IFR ontologies. As a consequence, the investigators are now positioned (starting in 2009, with a new award) to begin to build a general-purpose MCL module that, when attached to a given host program H and an initial set of IFR ontologies, can adapt to the domain that H lives in (and in the process adapt the ontologies to better fit that domain) so that H+MCL guides itself to become less brittle.

Computational approaches to learning aspects of language typically reduce the problem to learning syntax alone, or learning a lexicon alone. These simplifications have led to disconnected solutions and some unreasonable assumptions about inputs to their algorithms. In this paper, we present an approach that exploits a grammar learning algorithm to learn its own alphabet, or lexicon. We present empirical results and categorize the successes and types of errors lexical acquisition approaches encounter.

Segmenting real-valued data, be it speech waveforms into words and phrases or temperature readings into environmental epochs, is a challenging, open problem. We introduce an unsupervised, domain-independent algorithm, RIPTIDE, that discovers segments in real-valued time series data while constructing a hierarchy of segments. Our top-down approach begins with a coarse approximation of the input data, finds segment boundaries, and recursively considers discovered segments with a finer resolution. We demonstrate the drawbacks of an existing segmentation algorithm and the multiresolution capabilities of a discretization method for time series.

This effort developed simulation, visualization and evaluation technology for air campaign planning. A simple war game called Capture the Flag was used to build and evaluate military Courses of Action based on the developed technology. The Capture the Flag system comprises a simulator, a control architecture, and a planner. The system provides graphical user interfaces for users to observe and participate in war games. A variety of visualizations, instrumentation, and other software support the execution and analysis of experiments. Dynamic maps were developed to visualize the results of hundreds of Monte Carlo simulation trials. This visualization technique, referred to as probabilistic dynamic mapping, relates dynamical properties of a war game to the probability of winning the game. A method for unsupervised clustering of univariate, categorical time series was also developed. This method is referred to as Bayesian Clustering by Dynamics.

This paper describes PERUSE, an unsupervised algorithm for finding recurring patterns in time series. It was initially developed and tested with sensor data from a mobile robot, i.e. noisy, real-valued, multivariate time series with variable intervals between observations. The pattern discovery problem is decomposed into two subproblems: (1) a supervised learning problem in which a teacher provides exemplars of patterns and labels time series according to whether they contain the patterns; (2) an unsupervised learning problem in which the time series are used to generate an approximation to the teacher. Experimental results show that PERUSE can discover patterns in audio data corresponding to recurring words in natural language utterances and patterns in the sensor data of a mobile robot corresponding to qualitatively distinct outcomes of taking actions.

Presents a methodology for learning complex dependencies in data based on streams of categorical time-series data. The streams representation is applicable in a variety of situations. A program's execution trace may be thought of as a stream. The various monitor readings of an intensive care unit may be thought of as concurrent streams. Our learning methodology, called 'dependency detection', examines one or more streams to characterize a recurring structure with a set of dependency rules. These dependency rules are useful not only as a description of how the data is structured, but as a means for predicting future stream states. Further, we describe a set of tools for program analysis that use dependency detection.

The objective of this STTR project is to develop an information management system to rapidly and accurately linking records of related information from web-based information sources. The sheer magnitude of information available online via the Internet has overwhelmed the ability of existing search tools to produce useful query responses. Current web-search techniques typically fail to correlate relevant documents that are identified in different ways, such as synonyms and acronyms (aliases). The challenge is to find an approach that can obtain highly accurate matches even when those documents do not share any obvious attributes with the query, and with minimal information requirement from the user. Latent Semantic Analysis (LSA) is a technique for identifying both semantically similar words and semantically similar documents. On the face of it, LSA should work well for the task of discovering aliases. That is, for a given word we can use LSA to produce a rank-ordered list of words that are semantically similar to it and aliases for the name should be high in this list. In this Phase I, we tested this conjecture empirically and found, surprisingly, that under a broad range of circumstances a straightforward application of LSA fails to rank the aliases highly. We then developed a two-stage algorithm that takes the output of LSA, creates a new set of pseudo-documents, and runs LSA again on these new documents. Empirical results show that this two-stage algorithm performs remarkably well in identifying aliases, even in those cases for which a single application of LSA fails miserably. University of Maryland (Baltimore County) is the research institute partner for this effort, under the direction of Professor Charles Nicholas and Tim Oates.

A robotic agent experiences a world of continuous multivariate sensations and chooses its actions from continuous action spaces. Unless the agent is able to successfully partition these into functionally similar classes, its ability to interact with the world will be extremely limited. We present a method whereby an unsupervised robotic agent learns to discriminate discrete actions out of its continuous action parameters. These actions are discriminated because they lead to qualitatively distinct outcomes in the robot's sensor space. Once found, these actions can be used by the robot as primitives for further exploration of its world. We present results gathered using a Pioneer 1 mobile robot.

In this paper we claim that meaningful representations can be learned by programs, although today they are almost always designed by skilled engineers. We discuss several kinds of meaning that representations might have, and focus on a functional notion of meaning as appropriate for programs to learn. Specifically, a representation is meaningful if it incorporates an indicator of external conditions and if the indicator relation informs action. We survey methods for inducing kinds of representations we call structural abstractions. Prototypes of sensory time series are one kind of structural abstraction, and though they are not denoting or compositional, they do support planning. Deictic representations of objects and prototype representations of words enable a program to learn the denotational meanings of words. Finally, we discuss two algorithms designed to find the macroscopic structure of episodes in a domain-independent way.