Deep neural networks (DNNs) are becoming commonplace in critical applications, making their susceptibility to backdoor (trojan) attacks a significant problem. In this paper, we introduce a novel backdoor attack detection pipeline, detecting attacked models using graph convolution networks (DeBUGCN). To the best of our knowledge, ours is the first use of GCNs for trojan detection. We use the static weights of a DNN to create a graph structure of its layers. A GCN is then used as a binary classifier on these graphs, yielding a trojan or clean determination for the DNN. To demonstrate the efficacy of our pipeline, we train hundreds of clean and trojaned CNN models on the MNIST handwritten digits and CIFAR-10 image datasets, and show the DNN classification results using DeBUGCN. For a true In-the-Wild use case, our pipeline is evaluated on the TrojAI dataset which consists of various CNN architectures, thus showing the robustness and model-agnostic behaviour of DeBUGCN. Furthermore, on comparing our results on several datasets with state-of-the-art trojan detection algorithms, DeBUGCN is faster and more accurate.

Malware detection is an interesting and valuable domain to work in because it has significant real-world impact and unique machine-learning challenges. We investigate existing long-range techniques and benchmarks and find that they’re not very suitable in this problem area. In this paper, we introduce Holographic Global Convolutional Networks (HGConv) that utilize the properties of Holographic Reduced Representations (HRR) to encode and decode features from sequence elements. Unlike other global convolutional methods, our method does not require any intricate kernel computation or crafted kernel design. HGConv kernels are defined as simple parameters learned through backpropagation. The proposed method has achieved new SOTA results on Microsoft Malware Classification Challenge, Drebin, and EMBER malware benchmarks. With log-linear complexity in sequence length, the empirical results demonstrate substantially faster run-time by HGConv compared to other methods achieving far more efficient scaling even with sequence length $\geq 100,000$.

As deep neural networks and the datasets used to train them get larger, the default approach to integrating them into research and commercial projects is to download a pre-trained model and fine tune it. But these models can have uncertain provenance, opening up the possibility that they embed hidden malicious behavior such as trojans or backdoors, where small changes to an input (triggers) can cause the model to produce incorrect outputs (e.g., to misclassify). This paper introduces a novel approach to backdoor detection that uses two tensor decomposition methods applied to network activations. This has a number of advantages relative to existing detection methods, including the ability to analyze multiple models at the same time, working across a wide variety of network architectures, making no assumptions about the nature of triggers used to alter network behavior, and being computationally efficient. We provide a detailed description of the detection pipeline along with results on models trained on the MNIST digit dataset, CIFAR-10 dataset, and two difficult datasets from NIST’s TrojAI competition. These results show that our method detects backdoored networks more accurately and efficiently than current state-of-the-art methods.

In the rapidly evolving landscape of communication and network security, the increasing reliance on deep neural networks (DNNs) and cloud services for data processing presents a significant vulnerability: the potential for backdoors that can be exploited by malicious actors. Our approach leverages advanced tensor decomposition algorithms-Independent Vector Analysis (IVA), Multiset Canonical Correlation Analysis (MCCA), and Parallel Factor Analysis (PARAFAC2)-to meticulously analyze the weights of pre-trained DNNs and distinguish between back-doored and clean models effectively. The key strengths of our method lie in its domain independence, adaptability to various network architectures, and ability to operate without access to the training data of the scrutinized models. This not only ensures versatility across different application scenarios but also addresses the challenge of identifying backdoors without prior knowledge of the specific triggers employed to alter network behavior. We have applied our detection pipeline to three distinct computer vision datasets, encompassing both image classification and object detection tasks. The results demonstrate a marked improvement in both accuracy and efficiency over existing back-door detection methods. This advancement enhances the security of deep learning and AI in networked systems, providing essential cybersecurity against evolving threats in emerging technologies.

Vector Symbolic Architectures (VSAs) are one approach to developing Neuro-symbolic AI, where two vectors in $\mathbb{R}^d$ are 'bound' together to produce a new vector in the same space. VSAs support the commutativity and associativity of this binding operation, along with an inverse operation, allowing one to construct symbolic-style manipulations over real-valued vectors. Most VSAs were developed before deep learning and automatic differentiation became popular and instead focused on efficacy in hand-designed systems. In this work, we introduce the Hadamard-derived linear Binding (HLB), which is designed to have favorable computational efficiency, and efficacy in classic VSA tasks, and perform well in differentiable systems.

Learning High-Resolution representations is essential for semantic segmentation. Convolutional neural network (CNN) architectures with downstream and upstream propagation flow are popular for segmentation in medical diagnosis. However, due to performing spatial downsampling and upsampling in multiple stages, information loss is inexorable. On the contrary, connecting layers densely on high spatial resolution is computationally expensive. In this work, we devise a Loose Dense Connection Strategy to connect neurons in subsequent layers with reduced parameters. On top of that, using a m-way Tree structure for feature propagation we propose Receptive Field Chain Network (RFC-Net) that learns high-resolution global features on a compressed computational space. Our experiments demonstrates that RFC Net achieves state-of-the-art performance on Kvasir and CVC-ClinicDB benchmarks for Polyp segmentation. Our code is publicly available at github.com/sourajitcs/RFC-NetAAAI23.

In recent years, self-attention has become the dominant paradigm for sequence modeling in a variety of domains. However, in domains with very long sequence lengths the $\mathcal{O}(T^2)$ memory and $\mathcal{O}(T^2 H)$ compute costs can make using transformers infeasible. Motivated by problems in malware detection, where sequence lengths of $T \geq 100,000$ are a roadblock to deep learning, we re-cast self-attention using the neuro-symbolic approach of Holographic Reduced Representations (HRR). In doing so we perform the same high-level strategy of the standard self-attention: a set of queries matching against a set of keys, and returning a weighted response of the values for each key. Implemented as a ``Hrrformer'' we obtain several benefits including $\mathcal{O}(T H \log H)$ time complexity, $\mathcal{O}(T H)$ space complexity, and convergence in $10\times$ fewer epochs. Nevertheless, the Hrrformer achieves near state-of-the-art accuracy on LRA benchmarks and we are able to learn with just a single layer. Combined, these benefits make our Hrrformer the first viable Transformer for such long malware classification sequences and up to $280\times$ faster to train on the Long Range Arena benchmark.

Due to the computational cost of running inference for a neural network, the need to deploy the inferential steps on a third party’s compute environment or hardware is common. If the third party is not fully trusted, it is desirable to obfuscate the nature of the inputs and outputs, so that the third party can not easily determine what specific task is being performed. Provably secure protocols for leveraging an untrusted party exist but are too computational demanding to run in practice. We instead explore a different strategy of fast, heuristic security that we call Connectionist Symbolic Pseudo Secrets. By leveraging Holographic Reduced Representations (HRRs), we create a neural network with a pseudo-encryption style defense that empirically shows robustness to attack, even under threat models that unrealistically favor the adversary.

Recently, Reinforcement Learning (RL) has shown great performance in solving sequential decision-making and control in dynamic environment problems. Despite its achievements, deploying Deep Neural Network (DNN)-based RL is expensive in terms of time and power due to the large number of episodes required to train agents with high dimensional image representations. Additionally, at the interference the large energy footprint of deep neural networks can be a major drawback. Embedded edge devices as the main platform for deploying RL applications are intrinsically resource-constrained and deploying deep neural network-based RL on them is a challenging task. As a result, reducing the number of actions taken by the RL agent to learn desired policy, along with the energy-efficient deployment of RL, is crucial. In this article, we propose Energy Efficient Hierarchical Reinforcement Learning (E2HRL), which is a scalable hardware architecture for RL applications. E2HRL utilizes a cross-layer design methodology for achieving better energy efficiency, smaller model size, higher accuracy, and system integration at the software and hardware layers. Our proposed model for RL agent is designed based on the learning hierarchical policies, which makes the network architecture more efficient for implementation on mobile devices. We evaluated our model in three different RL environments with different level of complexity. Simulation results with our analysis illustrate that hierarchical policy learning with several levels of control improves RL agents training efficiency and the agent learns the desired policy faster compared to a non-hierarchical model. This improvement is specifically more observable as the environment or the task becomes more complex with multiple objective subgoals. We tested our model with different hyperparameters to achieve the maximum reward by the RL agent while minimizing the model size, parameters, and required number of operations. E2HRL model enables efficient deployment of RL agent on resource-constraint-embedded devices with the proposed custom hardware architecture that is scalable and fully parameterized with respect to the number of input channels, filter size, and depth. The number of processing engines (PE) in the proposed hardware can vary between 1 to 8, which provides the flexibility of tradeoff of different factors such as latency, throughput, power, and energy efficiency. By performing a systematic hardware parameter analysis and design space exploration, we implemented the most energy-efficient hardware architectures of E2HRL on Xilinx Artix-7 FPGA and NVIDIA Jetson TX2. Comparing the implementation results shows Jetson TX2 boards achieve 0.1 ∼ 1.3 GOP/S/W energy efficiency while Artix-7 FPGA achieves 1.1 ∼ 11.4 GOP/S/W, which denotes 8.8× ∼ 11× better energy efficiency of E2HRL when model is implemented on FPGA. Additionally, compared to similar works our design shows better performance and energy efficiency.

The Uniform Manifold Approximation and Projection (UMAP) algorithm has become widely popular for its ease of use, quality of results, and support for exploratory, unsupervised, supervised, and semi-supervised learning. While many algorithms can be ported to a GPU in a simple and direct fashion, such efforts have resulted in inefficent and inaccurate versions of UMAP. We show a number of techniques that can be used to make a faster and more faithful GPU version of UMAP, and obtain speedups of up to 100x in practice. Many of these design choices/lessons are general purpose and may inform the conversion of other graph and manifold learning algorithms to use GPUs. Our implementation has been made publicly available as part of the open source RAPIDS cuML library (https://github.com/rapidsai/cuml).

Holographic Reduced Representations (HRR) are a method for performing symbolic AI on top of real-valued vectors by associating each vector with an abstract concept, and providing mathematical operations to manipulate vectors as if they were classic symbolic objects. This method has seen little use outside of older symbolic AI work and cognitive science. Our goal is to revisit this approach to understand if it is viable for enabling a hybrid neural-symbolic approach to learning as a differential component of a deep learning architecture. HRRs today are not effective in a differential solution due to numerical instability, a problem we solve by introducing a projection step that forces the vectors to exist in a well behaved point in space. In doing so we improve the concept retrieval efficacy of HRRs by over $100\times$. Using multi-label classification we demonstrate how to leverage the symbolic HRR properties to develop a output layer and loss function that is able to learn effectively, and allows us to investigate some of the pros and cons of an HRR neuro-symbolic learning approach.

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 4-5X improved performance compared to a baseline RL agent with a convolutional feature extractor, while using less than 2W of power.

Task sharing between heterogeneous robots currently requires a priori capability knowledge, a shared communication protocol, or a centralized planner. However, in practice, when two robots are brought together, the effort required to construct shared action and structure models can be significant. In this paper, we describe our approach to determining the kinematic model of a robot based purely on observation of unscripted movement. We describe construction of large-scale data simulating low-cost RGB-D camera output, and application of two different RNN-based methods to the learning problem. Our results suggest that this is an efficient and effective way to determine a robot's morphological structure without requiring communication or pre-existing knowledge of its capabilities.

The problems of recurrent and anomalous pattern discovery in time series, e.g., motifs and discords, respectively, have received a lot of attention from researchers in the past decade. However, since the pattern search space is usually intractable, most existing detection algorithms require that the patterns have discriminative characteristics and have its length known in advance and provided as input, which is an unreasonable requirement for many real-world problems. In addition, patterns of similar structure, but of different lengths may co-exist in a time series. Addressing these issues, we have developed algorithms for variable-length time series pattern discovery that are based on symbolic discretization and grammar inference—two techniques whose combination enables the structured reduction of the search space and discovery of the candidate patterns in linear time. In this work, we present GrammarViz 3.0—a software package that provides implementations of proposed algorithms and graphical user interface for interactive variable-length time series pattern discovery. The current version of the software provides an alternative grammar inference algorithm that improves the time series motif discovery workflow, and introduces an experimental procedure for automated discretization parameter selection that builds upon the minimum cardinality maximum cover principle and aids the time series recurrent and anomalous pattern discovery.

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. 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. This translates to real-world control problems such as those in robotics and automation (e.g. the robust output tracking problem or controlling an n-joint arm), where the underlying equations can be made linear. A conclusive performance hierarchy between evaluated algorithms is presented.

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.

Neural networks have attracted significant interest in recent years due to their exceptional performance in various domains ranging from natural language processing to image identification and classification. Modern deep neural networks demonstrate state-of-the-art results in complex tasks such as epileptic seizure detection [1] and time series classification [2]. The internal architecture of these networks, in terms of learned representations, still remains opaque. This research addresses the first step towards the long term goal of constructing a bidirectional connection between raw input data and symbolic representations. In this research, we examined whether a denoising autoencoder can internally find correlated principal features from input images and their symbolic representations that can be used to generate one from the other. Our results indicate that using symbolic representations along with the raw inputs generates better reconstructions. Our network was able to construct the symbolic representations from the input as well as input instances from their symbolic representations.

Entity linking is the task of identifying entities like people and places in textual data and linking them to corresponding entities in a knowledge base. In this paper we solve a visual equivalent of this task called visual entity linking. The goal is to link regions of images to corresponding entities in knowledge bases. Visual entity linking will enable computers to better understand visual content and thus can be used in tasks like image retrieval and visual question answering. More specifically, we propose a novel approach for linking image regions to entities in Dbpedia and Freebase. First, we select candidate entities using an automatic image description generation algorithm. We then extract image regions using object detection methods and compare them to depictions of entities in a knowledge base. We evaluate our approach on the Flickr8k dataset through surveys on Amazon Mechanical Turk, and present an extensive analysis to identify the sources of errors in our system.

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.

Electroencephalogram (EEG) data is used for a variety of purposes, including brain-computer interfaces, disease diagnosis, and determining cognitive states. Yet EEG signals are susceptible to noise from many sources, such as muscle and eye movements, and motion of electrodes and cables. Traditional approaches to this problem involve supervised training to identify signal components corresponding to noise so that they can be removed. However these approaches are artifact specific. In this paper, we present a novel software-hardware system that uses a weak supervisory signal to indicate that some noise is occurring, but not what the source of the noise is or how it is manifested in the EEG signal. The EEG data is decomposed into independent components using ICA, and these components form bags that are labeled and classified by a multi-instance learning algorithm that can identify the noise components for removal to reconstruct a clean EEG signal. We also performed extensive hyperparameter optimization for the model with the goal of improving accuracy without increasing execution time. This resulted the execution time to be reduced from 282 s to 8.8 s when running the model on an embedded ARM CPU processor at 1.6 GHz clock frequency. In this paper, we present the overall system which includes ICA, SAX and MIL, along with preliminary results for software and hardware implementation when using real EEG data from 64 electrodes. The proposed system consumes 909 mW power during processing above a baseline of 2.32 W idle, while achieving 91.2% artifact identification accuracy.

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.

This paper proposes a set of new error criteria and a learning approach, called Adaptive Normalized Risk-Averting Training (ANRAT) to attack the non-convex optimization problem in training deep neural networks without pretraining. Theoretically, we demonstrate its effectiveness based on the expansion of the convexity region. By analyzing the gradient on the convexity index $\lambda$, we explain the reason why our learning method using gradient descent works. In practice, we show how this training method is successfully applied for improved training of deep neural networks to solve visual recognition tasks on the MNIST and CIFAR-10 datasets. Using simple experimental settings without pretraining and other tricks, we obtain results comparable or superior to those reported in recent literature on the same tasks using standard ConvNets + MSE/cross entropy. Performance on deep/shallow multilayer perceptron and Denoised Auto-encoder is also explored. ANRAT can be combined with other quasi-Newton training methods, innovative network variants, regularization techniques and other common tricks in DNNs. Other than unsupervised pretraining, it provides a new perspective to address the non-convex optimization strategy in training DNNs.

Time series classication is an important problem that has received a great amount of attention by researchers and practitioners in the past two decades. In this work, we propose a novel algorithm for time series classication based on the discovery of class-specic representative patterns. We dene representative patterns of a class as a set of subsequences that has the greatest discriminative power to distinguish one class of time series from another. Our approach rests upon two techniques with linear complexity: symbolic discretization of time series, which generalizes the structural patterns, and grammatical inference, which automatically nds recurrent correlated patterns of variable length, producing an initial pool of common patterns shared by many instances in a class. From this pool of candidate patterns, our algorithm selects the most representative patterns that capture the class specicities, and that can be used to eectively discriminate between time series classes. Through an exhaustive experimental evaluation we show that our algorithm is competitive in accuracy and speed with the stateof-the-art classication techniques on the UCR time series repository, robust on shifted data, and demonstrates excellent performance on real-world noisy medical time series.

Though graph grammars have been widely investigated for 40 years, few learning results exist for them. The main reasons come from complexity issues that are inherent when graphs, and a fortiori graph grammars, are considered. The picture is however different if one considers drawings of graphs, rat her than the graphs themselves. For instance, it has recently been proved that the isomorphism and pattern searching problems could be solved in polynomial time for plane graphs, that is, planar embeddings of planar graphs. In this paper, we introduce the Plane Graph Grammars (PGG) and detail how they differ to usual graph grammar formalism’s while at the same time they share important properties with string context-free grammars. In particular, though exponential in the general case, we provide an appropriate restriction on languages that allows the parsing of a graph with a given PGG in polynomial time. We demonstrate that PGG are well-shaped for learning: we show how recent results on string grammars can be extended to PGG by providing a learning algorithm that identifies in the limit the class of substitutable plane graph languages. Our algorithm runs in polynomial time assuming the same restriction used for polynomial parsing, and the amount of data needed for convergence is comparable to the one required in the case of strings.

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.

Standard Symbolic Aggregation Approximation (SAX) is at the core of many effective time series data mining algorithms. Its combination with Bag-of-Patterns (BoP) has become the standard approach with state-of-the-art performance on standard datasets. However, standard SAX with the BoP representation might neglect internal temporal correlation embedded in the raw data. In this paper, we proposed time warping SAX, which extends the standard SAX with time delay embedding vector approaches to account for temporal correlations. We test time warping SAX with the BoP representation on 12 benchmark datasets from the UCR Time Series Classification/Clustering Collection. On 9 datasets, time warping SAX overtakes the state-of-the-art performance of the standard SAX. To validate our methods in real world applications, a new dataset of vital signs data collected from patients who may require blood transfusion in the next 6 hours was tested. All the results demonstrate that, by considering the temporal internal correlation, time warping SAX combined with BoP improves classification performance.

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.

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.

Monitoring of non-invasive, continuous, high-resolution patient vital signs (VS) such as heart rate and oxygen saturation is becoming increasingly common in hospital settings. These data are a potential boon for health informatics as a source of predictive information about a variety of patient outcomes. Yet the volume, noisiness, and per-patient idiosyncrasies of these data make their use extremely challenging. This paper explores the utility of representing VS data as unordered collections (bags) of local discrete patterns for the purpose of training classifiers to predict outcomes for traumatic brain injury patients, including mortality and level of cognitive function months after hospital discharge. The Symbolic Aggregate approXimation (SAX) algorithm is used for discretization, producing a bag of SAX "words" (local patterns) for each time series. Experiments with a dataset of sixty traumatic brain injury patients demonstrate that this approach is promising both in terms of predictive accuracy and patterns that it can reveal in the underlying VS data.

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.

The problem of time series motif discovery has received a lot of attention from researchers in the past decade. Most existing work on finding time series motifs require that the length of the motifs be known in advance. However, such information is not always available. In addition, motifs of different lengths may co-exist in a time series dataset. In this work, we develop a motif visualization system based on grammar induction. We demonstrate that grammar induction in time series can effectively identify repeated patterns without prior knowledge of their lengths. The motifs discovered by the visualization system are variable-lengths in two ways. Not only can the inter-motif subsequences have variable lengths, the intra-motif subsequences also are not restricted to have identical length—a unique property that is desirable, but has not been seen in the literature.

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.

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.

Sensor networks have gained importance in tracking of vital information. The accuracy of the sensor readings plays a major role in crucial applications like military and defense operations, and environment monitoring where the error margin has to be extremely small. In addition, sensor nodes are typically constrained in their computation, communication, and energy capacities and there is usually a need for managing sensor activities to utilize these limited resources in a judicious manner. Therefore, a participation criterion for the sensor nodes to balance fidelity of the collected data and conservation of nodes resources is important. This paper presents a novel node engagement strategy. The idea is to convoy a subset of the collected data reports while ensuring consensus of the nodes. Basically, nodes rotate transmission duties among neighbors. The data sent to the base-station will be based on the readings of the designated sensor. The neighboring nodes will act as passive listeners and stay quiet if the data does not deviate significantly from their measurements, which implies their endorsement of the accuracy of what the base-station will get. Otherwise, a subset of these nodes comes forward and transmits their readings so that the base-station can aggregate and accurately estimate the true value. Simulation results confirm the effectiveness of the proposed approach in terms of energy conservation and responsiveness to drops in data fidelity.

Many important domains are naturally described relationally, often using graphs in which nodes correspond to entities and edges to relations. Stochastic graph grammars compactly represent probability distributions over graphs and can be learned from data, such as a set of graphs corresponding to proteins that have the same function. In this paper we show that a stochastic graph grammar can be used to efficiently compute the probability mass functions of the number of nodes, the number of edges, and the degree of a node selected uniformly at random from a graph sampled from the distribution defined by the graph grammar. Both the expectation and variance of these quantities can be computed from the mass functions. Empirical results with standard synthetic grammars and a grammar from the domain of AIDS research demonstrate the accuracy of our methods.

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.

Many real-world domains resist analysis because they are best characterized by a variety of data types, including relational, spatial, and temporal components. Examples of such domains include disease outbreaks, criminal networks, and the World-Wide Web. We present two types of visualizations based on physical metaphors that facilitate fusion, analysis, and deep understanding of relational, spatio-temporal data. The first visualization is based on the metaphor of fluid flow through elastic pipes, and the second on wave propagation. We discuss both types of visualizations in the context of fusing information about the activities of scientists over time with the goal of constructing career histories

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.

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.

Most time series comparison algorithms attempt to discover what the members of a set of time series have in common. We investigate a different problem, determining what distinguishes time series in that set from other time series obtained from the same source. In both cases the goal is to identify shared patterns, though in the latter case those patterns must be distinctiveaswell. An efficient incremental algorithm for identifying distinctive subsequences in multivariate, real-valued time series is described and evaluated with data from two very different sources: the response of a set of bandpass filters to human speech and the sensors of a mobile robot.

Having access to massive amounts of data does not necessarily imply that induction algorithms must use them all.Samples often provide the same accuracy with far less computational cost.However, the correct sample size rarely is obvious.We analyze methods for progressive samplingusing progressively larger samples as long as model accuracy improves.We explore several notions of efficient progressive sampling.We analyze efficiency relative to induction with all instances; we show that a simple, geometric sampling schedule is asymptotically optimal, and we describe how best to take into account prior expectations of accuracy convergence.We then describe the issues involved in instantiating an efficient progressive sampler, including how to detect convergence.Finally, we provide empirical results comparing a variety of progressive sampling methods.We conclude that progressive sampling can be remarkably efficient .

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.