Eric Granger

A survey of techniques for incremental learning of HMM parameters

The performance of Hidden Markov Models (HMMs) targeted for complex real-world applications are often degraded because they are designed a priori using limited training data and prior knowledge, and because the classification environment changes during operations. Incremental learning of new data sequences allows to adapt HMM parameters as new data becomes available, without having to retrain from the start on all accumulated training data. This paper presents a survey of techniques found in literature that are suitable for incremental learning of HMM parameters. These techniques are classified according to the objective function, optimization technique and target application, involving block-wise and symbol-wise learning of parameters. Convergence properties of these techniques are presented along with an analysis of time and memory complexity. In addition, the challenges faced when these techniques are applied to incremental learning is assessed for scenarios in which the new training data is limited and abundant. While the convergence rate and resource requirements are critical factors when incremental learning is performed through one pass over abundant stream of data, effective stopping criteria and management of validation sets are important when learning is performed through several iterations over limited data. In both cases managing the learning rate to integrate pre-existing knowledge and new data is crucial for maintaining a high level of performance. Finally, this paper underscores the need for empirical benchmarking studies among techniques presented in literature, and proposes several evaluation criteria based on non-parametric statistical testing to facilitate the selection of techniques given a particular application domain.

A what-and-where fusion neural network for recognition and tracking of multiple radar emitters

A neural network recognition and tracking system is proposed for classification of radar pulses in autonomous Electronic Support Measure systems. Radar type information is considered with position-specific information from active emitters in a scene. Type-specific parameters of the input pulse stream are fed to a neural network classifier trained on samples of data collected in the field. Meanwhile, a clustering algorithm is used to separate pulses from different emitters according to position-specific parameters of the input pulse stream. Classifier responses corresponding to different emitters are separated into tracks, or trajectories, one per active emitter, allowing for more accurate identification of radar types based on multiple views of emitter data along each emitter trajectory. Such a What-and-Where fusion strategy is motivated by a similar subdivision of labor in the brain. The fuzzy ARTMAP neural network is used to classify streams of pulses according to radar type using their functional parameters. Simulation results obtained with a radar pulse data set indicate that fuzzy ARTMAP compares favorably to several other approaches when performance is measured in terms of accuracy and computational complexity. Incorporation into fuzzy ARTMAP of negative match tracking (from ARTMAP-IC) facilitated convergence during training with this data set. Other modifications improved classification of data that include missing input pattern components and missing training classes. Fuzzy ARTMAP was combined with a bank of Kalman filters to group pulses transmitted from different emitters based on their position-specific parameters, and with a module to accumulate evidence from fuzzy ARTMAP responses corresponding to the track defined for each emitter. Simulation results demonstrate that the system provides a high level of performance on complex, incomplete and overlapping radar data.

An adaptive classification system for video-based face recognition

In many practical applications, new information may emerge from the environment at different points in time after a classification system has originally been deployed. For instance, in biometric systems, new data may be acquired and used to enroll or to update knowledge of an individual. In this paper, an adaptive classification system (ACS) is proposed for video-based face recognition. It combines a fuzzy ARTMAP neural network classifier, dynamic particle swarm optimization (DPSO) algorithm, and a long term memory (LTM). A novel DPSO-based learning strategy is also presented for incremental learning of new data with this ACS. This strategy allows to cojointly optimize the classifier weights, architecture, and user-defined hyperparameters such as classification rate is maximized. Performance of this system is assessed in terms of classification rate and resource requirements for incremental learning of data blocks coming from real-world video data bases. The necessity of a LTM to store validation data is shown empirically for different enrollment and update scenarios. In addition, incremental learning is shown to constitute a dynamic optimization problem where the optimal hyperparameter values change in time. Simulation results indicate that the proposed system can provide a significant higher classification rate than that of fuzzy ARTMAP alone during incremental learning. However, optimization of ACS parameters requires more resources. The ACS needs several training sequences to produce the optimal solution, and adapting fuzzy ARTMAP parameters according to classification rate tends to require more category neurons and training epochs.

Dynamic selection of generative-discriminative ensembles for off-line signature verification

In practice, each writer provides only a limited number of signature samples to design a signature verification (SV) system. Hybrid generative-discriminative ensembles of classifiers (EoCs) are proposed in this paper to design an off-line SV system from few samples, where the classifier selection process is performed dynamically. To design the generative stage, multiple discrete left-to-right Hidden Markov Models (HMMs) are trained using a different number of states and codebook sizes, allowing the system to learn signatures at different levels of perception. To design the discriminative stage, HMM likelihoods are measured for each training signature, and assembled into feature vectors that are used to train a diversified pool of two-class classifiers through a specialized Random Subspace Method. During verification, a new dynamic selection strategy based on the K-nearest-oracles (KNORA) algorithm and on Output Profiles selects the most accurate EoCs to classify a given input signature. This SV system is suitable for incremental learning of new signature samples. Experiments performed with real-world signature data (composed of genuine samples, and random, simple and skilled forgeries) indicate that the proposed dynamic selection strategy can significantly reduce the overall error rates, with respect to other EoCs formed using well-known dynamic and static selection strategies. Moreover, the performance of the SV system proposed in this paper is significantly greater than or comparable to that of related systems found in the literature.

Multi-feature extraction and selection in writer-independent off-line signature verification

Some of the fundamental problems faced in the design of signature verification (SV) systems include the potentially large number of input features and users, the limited number of reference signatures for training, the high intra-personal variability among signatures, and the lack of forgeries as counterexamples. In this paper, a new approach for feature selection is proposed for writer-independent (WI) off-line SV. First, one or more preexisting techniques are employed to extract features at different scales. Multiple feature extraction increases the diversity of information produced from signature images, allowing to produce signature representations that mitigate intra-personal variability. Dichotomy transformation is then applied in the resulting feature space to allow for WI classification. This alleviates the challenges of designing off-line SV systems with a limited number of reference signatures from a large number of users. Finally, boosting feature selection is used to design low-cost classifiers that automatically select relevant features while training. Using this global WI feature selection approach allows to explore and select from large feature sets based on knowledge of a population of users. Experiments performed with real-world SV data comprised of random, simple, and skilled forgeries indicate that the proposed approach provides a high level of performance when extended shadow code and directional probability density function features are extracted at multiple scales. Comparing simulation results to those of off-line SV systems found in literature confirms the viability of the new approach, even when few reference signatures are available. Moreover, it provides an efficient framework for designing a wide range of biometric systems from limited samples with few or no counterexamples, but where new training samples emerge during operations.

Classification of incomplete data using the fuzzy ARTMAP neural network

The fuzzy ARTMAP neural network is used to classify data that is incomplete in one or more ways. These include a limited number of training cases, missing components, missing class labels, and missing classes. Modifications for dealing with such incomplete data are introduced, and performance is assessed on an emitter identification task using a database of radar pulses.

Adaptive ROC-based ensembles of HMMs applied to anomaly detection

Hidden Markov models (HMMs) have been successfully applied in many intrusion detection applications, including anomaly detection from sequences of operating system calls. In practice, anomaly detection systems (ADSs) based on HMMs typically generate false alarms because they are designed using limited amount of representative training data. Since new data may become available over time, an important feature of an ADS is the ability to accommodate newly acquired data incrementally, after it has originally been trained and deployed for operations. In this paper, a system based on the receiver operating characteristic (ROC) is proposed to efficiently adapt ensembles of HMMs (EoHMMs) in response to new data, according to a learn-and-combine approach. When a new block of training data becomes available, a pool of base HMMs is generated from the data using a different number of HMM states and random initializations. The responses from the newly trained HMMs are then combined to those of the previously trained HMMs in ROC space using a novel incremental Boolean combination (incrBC) technique. Finally, specialized algorithms for model management allow to select a diversified EoHMM from the pool, and adapt Boolean fusion functions and thresholds for improved performance, while it prunes redundant base HMMs. The proposed system is capable of changing the desired operating point during operations, and this point can be adjusted to changes in prior probabilities and costs of errors. Computer simulations conducted on synthetic and real-world host-based intrusion detection data indicate that the proposed system can achieve a significantly higher level of performance than when parameters of a single best HMM are estimated, at each learning stage, using reference batch and incremental learning techniques. It also outperforms the learn-and-combine approaches using static fusion functions (e.g., majority voting). Over time, the proposed ensemble selection algorithms form compact EoHMMs, while maintaining or improving system accuracy. Pruning allows to limit the pool size from increasing indefinitely, thereby reducing the storage space for accommodating HMMs parameters without negatively affecting the overall EoHMM performance. Although applied for HMM-based ADSs, the proposed approach is general and can be employed for a wide range of classifiers and detection applications.

Supervised learning of fuzzy ARTMAP neural networks through particle swarm optimization

In this paper, the impact on fuzzy ARTMAP performance of decisions taken for batch supervised learning is assessed through computer simulation. By learning different realworld and synthetic data, using different learning strategies, training set sizes, and hyperparameter values, the generalization error and resources requirements of this neural network are compared. In particular, the degradation of fuzzy ARTMAP performance due to overtraining is shown to depend on factors such as the training set size and the number of training epochs, and occur for pattern recognition problems in which class distributions overlap. Although the hold-out learning strategy is commonly employed to avoid overtraining, results indicate that it is not necessarily justified. As an alternative, a new Particle Swarm Optimization (PSO) learning strategy, based on the concept of neural network evolution, has been introduced. It co-jointly determines the weights, architecture and hyper-parameters such that generalization error is minimized. Through a comprehensive set of simulations, it has been shown that when fuzzy ARTMAP uses this strategy, it produces a significantly lower generalization error, and mitigates the degradation of error due to overtraining. Overall, the results reveal the importance of optimizing all fuzzy ARTMAP parameters for a given problem, using a consistent objective function.

On the memory complexity of the forward-backward algorithm

The Forward-backward (FB) algorithm forms the basis for estimation of Hidden Markov Model (HMM) parameters using the Baum-Welch technique. It is however, known to be prohibitively costly when estimation is performed from long observation sequences. Several alternatives have been proposed in literature to reduce the memory complexity of FB at the expense of increased time complexity. In this paper, a novel variation of the FB algorithm - called the Efficient Forward Filtering Backward Smoothing (EFFBS) - is proposed to reduce the memory complexity without the computational overhead. Given an HMM with N states and an observation sequence of length T, both FB and EFFBS algorithms have the same time complexity, O(N^2T). Nevertheless, FB has a memory complexity of O(NT), while EFFBS has a memory complexity that is independent of T, O(N). EFFBS requires fewer resources than FB, yet provides the same results.

Hybrid writer-independent–writer-dependent offline signature verification system

Standard signature verification (SV) systems are writer-dependent (WD), where a specific classifier is designed for each individual. It is inconvenient to ask a user to provide enough number of signature samples to design his WD classifier. In practice, very few samples are collected and inaccurate classifiers maybe produced. To overcome this, writer-independent (WI) systems are introduced. A global classifier is designed using a development database, prior to enrolling users to the system. For these systems, signature templates are needed for verification, and the template databases can be compromised. Moreover, state-of-the-art WI and WD systems provide enhanced accuracy through information fusion at either feature, score or decision levels, but they increase computational complexity. In this study, a hybrid WI-WD system is proposed, as a compromise of the two approaches. When a user is enrolled to the system, a WI classifier is used to verify his queries. During operation, user samples are collected and adapt the WI classifier to his signatures. Once adapted, the resulting WD classifier replaces the WI classifier for this user. Simulations on the Brazilian and the GPDS signature databases indicate that the proposed hybrid system provides comparative accuracy as complex WI and WD systems, while decreases the classification complexity.