20 December 2016 in IEEE Transactions on Neural Networks and Learning Systems
Localized multiple kernel learning (LMKL) is an attractive strategy for combining multiple heterogeneous features with regard to their discriminative power for each individual sample. However, the learning of numerous local solutions may not scale well even for a moderately sized training set, and the independently learned local models may suffer from overfitting. Hence, in existing local methods, the distributed samples are typically assumed to share the same weights, and various unsupervised clustering methods are applied as preprocessing. In this paper, to enable the learner to discover and benefit from the underlying local coherence and diversity of the samples, we incorporate the clustering procedure into the canonical support vector machine-based LMKL framework. Then, to explore the relatedness among different samples, which has been ignored in a vector $\ell _{p}$ -norm analysis, we organize the cluster-specific kernel weights into a matrix and introduce a matrix-based extension of the $\ell _{p}$ -norm for constraint enforcement. By casting the joint optimization problem as a problem of alternating optimization, we show how the cluster structure is gradually revealed and how the matrix-regularized kernel weights are obtained. A theoretical analysis of such a regularizer is performed using a Rademacher complexity bound, and complementary empirical experiments on real-world data sets demonstrate the effectiveness of our technique.
15 January 2018 in IEEE Transactions on Neural Networks and Learning Systems
This paper examines a matrix-regularized multiple kernel learning (MKL) technique based on a notion of (r, p) norms. For the problem of learning a linear combination in the support vector machine-based framework, model complexity is typically controlled using various regularization strategies on the combined kernel weights. Recent research has developed a generalized ℓ p -norm MKL framework with tunable variable p(p ≥ 1) to support controlled intrinsic sparsity. Unfortunately, this “1-D” vector ℓ p -norm hardly exploits potentially useful information on how the base kernels “interact.” To allow for higher order kernel-pair relationships, we extend the “1-D” vector ℓ p -MKL to the “2-D” matrix (r, p) norms (1 ≤ r, p <; ∞). We develop a new formulation and an efficient optimization strategy for (r, p)-MKL with guaranteed convergence. A theoretical analysis and experiments on seven UCI data sets shed light on the superiority of (r, p)-MKL over ℓ p -MKL in various scenarios.
11 April 2017 in IEEE Transactions on Neural Networks and Learning Systems
This brief analyzes the effects of regularization variations in the localized kernel weights on the hypothesis generated by localized multiple kernel learning (LMKL) algorithms. Recent research on LMKL includes imposing different regularizations on the localized kernel weights and has led to varying formulations and solution strategies. Following the stability analysis theory as presented by Bousquet and Elisseeff, we give stability bounds based on the norm of the variation of localized kernel weights for three LMKL methods cast in the support vector machine classification framework, including vector ℓ p -norm LMKL, matrix-regularized (r, p)-norm LMKL, and samplewise ℓ p -norm LMKL. Further comparison of these bounds helps to qualitatively reveal the performance differences produced by these regularization methods, that is, matrix-regularized LMKL achieves superior performance, followed by vector ℓ p -norm LMKL and samplewise ℓ p -norm LMKL. Finally, a set of experimental results on ten benchmark machine learning UCI data sets is reported and shown to empirically support our theoretical analysis.
1 November 2020 in The Journal of the Acoustical Society of America
High-order time lacunarity (HOT-Lac) is an effective feature for characterizing active sonar echographs of harbor environments. However, it involves high computational complexity of loop summations. Motivated by the idea of integral image, this Letter extends an echo-intensity integral sequence, a representation of filtering with time domain recursion, permitting fast and online updates of HOT-Lac in a constant number of operations. Evaluated by a set of real-world harbor data, the proposed method is capable of computing HOT-Lac extremely rapidly while maintaining equivalent area under curve performances to its off-line counterpart, demonstrating its potential for real-time surveillance of the harbor.
1 April 2023 in The Journal of the Acoustical Society of America
Active tracking of underwater small targets is a great challenge with kinematic information alone. This is because the active sonar often encounters multipath propagation and the induced clutter can even mask target echoes. Recently, high-order time lacunarity (HOT-Lac) has shown its ability in effectively highlighting “blob” targets from high clutter harbor environments. Hence, this paper proposes a HOT-Lac aided track scoring mechanism to solve the ambiguity of data association within the framework of Multiple Hypotheses Tracking (MHT). Specifically, the trajectory consistency of potential targets is captured by a momentum accumulation of the HOT Lac feature, which can inherit the historical information for the whole track. Meanwhile, due to the separability of the distribution of target and clutter in the HOT-Lac feature space, the probabilities of the target hypothesis and null hypothesis are modeled by the online computation of the HOT-Lac feature. Finally, the cumulative likelihood ratio based on HOT-Lac is integrated into MHT to score the potential tracks. Experiments in several real-world harbor scenarios demonstrate that the proposed HOT-Lac feature-aided tracker can suppress false tracks accurately and quickly.
1 May 2023 in The Journal of the Acoustical Society of America
An important goal of an active sonar system is to detect and track underwater intruders such as frogmen, unmanned underwater vehicles, etc. Unfortunately, the intruders appear visually as a small fluctuating “blob” against the high-level fluctuating background caused by multipath propagation and reverberation in the harbor environment, making it difficult to be distinguished. Classical motion features well developed in computer vision cannot cope with an underwater environment. Thus, this paper presents a robust high-order flux tensor (RHO-FT) to characterize the small underwater moving targets against high-level fluctuating background. According to the dynamic behavior of active clutter from real-world harbor environment, we first classify it into two main types: (1) dynamic clutter but with relatively consistent spatial-temporal variation in a certain neighborhood; (2) sparkle clutter presenting completely random flashing. Then starting from the classical flux tensor, we develop a statistical high-order computation to handle the former followed by a spatial-temporal connected component to suppress the latter to achieve higher robustness. Experiments on a set of real-world harbor datasets demonstrate the effectiveness of our RHO-FT.