Category: Neurocomputing

Slow Feature Analysis (SFA) is an unsupervised learning algorithm which extracts slowly varying features from a temporal vectorial signal. In SFA, feature slowness is measured by the average value of its squared time-derivative. In this paper, we introduce Frequency-Based Slow Feature Analysis (FSFA) and prove that it is a generalization of SFA in the frequency domain. In FSFA, the low pass filtered versions of the extracted slow features have maximum energy, making slowness a filter dependent measurement. Experimental results show that the extracted features depend on the selected filter kernel and are different than the signals extracted using SFA. However, it is proven that there is one filter kernel that makes FSFA equivalent to SFA. Furthermore, experiments on UCF-101 video action recognition dataset, showcase that the features extracted by FSFA, with proper filter kernels, result in improved classification performance when compared to the features extracted by standard SFA. Finally, an experiment on UCF-101, with an indicative, simple and shallow neural network, being composed of FSFA and SFA nodes, demonstrates that the previously mentioned network, can transform the features extracted by a known Convolutional Neural Network to a new feature space, where classification performance through Support Vector Machine can be improved.

It is an explicit mode to use the clicking points by the mouse in the interactive image segmentation, while an implicit interaction mode is to use the fixation points from the eye-tracking device. Both modes can provide a series of points. Inspired by the similarity between these two interaction modes, we propose a novel human visual system (HVS) based neural network for transferring the constrained fixation point based segmentation to the clicking point based interactive segmentation. Briefly speaking, the sequence of information transmission and processing in our model is RGB image, VGG-16 backbone, LGN-like module (LGNL) and ConvLSTM block, which correspond to the pathway of stimulus transmission and processing, i.e. stimulus, retina, lateral geniculate nucleus (LGN) and visual cortex in the HVS. First, the RGB image is fed to the VGG-16 backbone to obtain the multiple-layer feature maps. Then the LGNL is adopted to effectively incorporate edge-aware features and semantic features from different layers of the VGG-16 backbone in multiple resolutions, so as to produce rich contextual features. Finally, with the guidance of the fixation density map transformed from the fixation points, the output feature maps of LGNL are utilized to generate the segmentation map via a stack of ConvLSTM blocks in a coarse-to-fine manner. Comprehensive experiments demonstrate that the proposed HVS based neural network achieves a higher segmentation performance and outperforms seven state-of-the-art methods, and prove that the transfer from constrained fixation points to clicking points is reasonable and valid.

We propose a new online approach for multimodal dictionary learning. The method developed in this work addresses the great challenges posed by the computational resource constraints in dynamic environment when dealing with large scale tensor sequences. Given a sequence of tensors, i.e. a set composed of equal-size tensors, the approach proposed in this paper allows to infer a basis of latent factors that generate these tensors by sequentially processing a small number of data samples instead of using the whole sequence at once. Our technique is based on block coordinate descent, gradient descent and recursive computations of the gradient. A theoretical result is provided and numerical experiments on both real and synthetic data sets are performed.

Latent factor (LF) models are greatly efficient in extracting valuable knowledge from High-Dimensional and Sparse (HiDS) matrices which are usually seen in many industrial applications. Stochastic gradient descent (SGD) is an effective algorithm to build an LF model, yet its convergence rate depends vastly on the learning rate which should be tuned with care. Therefore, automatic selection of an optimal learning rate for an SGD-based LF model is a meaningful issue. To address it, this study incorporates the principle of particle swarm optimization (PSO) into an SGD-based LF model for searching an optimal learning rate automatically. With it, we further propose an adaptive Latent Factor (ALF) model. Empirical studies on four HiDS matrices from real industrial applications indicate that an ALF model obvious outperforms an LF model according to convergence rate, and maintain competitive prediction accuracy for missing data.

The problem of Markov decision processes (MDPs) with uncertain observation noise has rarely been studied. This paper proposes a Robust Quadratic Programming(RQP) approach to approximate Bellman equation solution. Besides efficiency, the proposed algorithm exhibits great robustness against uncertain observation noise, which is essential in real world applications. We further represent the solution into kernel forms, which implicitly expands the state-encoded feature space to higher or even infinite dimensions. Experimental results well justify its efficiency and robustness. The comparison with different kernels demonstrates its flexibility of kernel selection for different application scenarios.