Search

Abstract It is essential to delve into the strategy of multimodal model pre-training, which is an obvious impact on downstream tasks. Currently, clustering learning has achieved noteworthy benefits in multiple methods. However, due to the availability of open image-text pairs, it is challenging for multimodal with clustering learning. In this paper, we propose an approach that utilizes clustering swap prediction strategy to learn image-text clustering embedding space by interaction prediction between image and text features. Unlike existing models with clustering learning, our method (Clus) allows for an open number of clusters for web-scale alt-text data. Furthermore, in order to train the image and text encoders efficiently, we introduce distillation learning approach and evaluate the performance of the image-encoder in downstream visual tasks. In addition, Clus is pre-trained end-to-end by using large-scale image-text pairs. Specifically, both text and image serve as ground truth for swap prediction, enabling effective representation learning. Concurrently, extensive experiments demonstrate that Clus achieves state-of-the-art performance on multiple downstream fine-tuning and zero-shot tasks (i.e., Image-Text Retrieval, VQA, NLVR2, Image Captioning, Object Detection, and Semantic Segmentation).

It is essential to delve into the strategy of multimodal model pre-training, which is an obvious impact on downstream tasks. Currently, clustering learning has achieved noteworthy benefits in multiple methods. However, due to the availability of open image-text pairs, it is challenging for multimodal with clustering learning. In this paper, we propose an approach that utilizes clustering swap prediction strategy to learn image-text clustering embedding space by interaction prediction between image and text features. Unlike existing models with clustering learning, our method (Clus) allows for an open number of clusters for web-scale alt-text data. Furthermore, in order to train the image and text encoders efficiently, we introduce distillation learning approach and evaluate the performance of the image-encoder in downstream visual tasks. In addition, Clus is pre-trained end-to-end by using large-scale image-text pairs. Specifically, both text and image serve as ground truth for swap prediction, enabling effective representation learning. Concurrently, extensive experiments demonstrate that Clus achieves state-of-the-art performance on multiple downstream fine-tuning and zero-shot tasks (i.e., Image-Text Retrieval, VQA, NLVR2, Image Captioning, Object Detection, and Semantic Segmentation). [ABSTRACT FROM AUTHOR]

Copyright of Erzincan University Journal of Science & Technology is the property of Erzincan Binali Yildirim Universitesi and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)

Despite marked improvements in tropical cyclone (TC) track ensemble forecasting, forecasters still have difficulty in making quick decisions when facing multiple potential predictions, so it is demanding to develop post-processing techniques reducing the uncertainty in TC track forecasts, and one of such techniques is the cluster-based methods. To improve the effect and efficiency of the previous cluster-based methods, this study adopts recombination clustering (RC) by optimizing the use of limited TC variables and constructing better features that can accurately capture the good TC track forecasts from the ensemble prediction system (EPS) of the China Meteorological Administration Tropical Regional Atmosphere Model for the South China Sea (CMA-TRAMS). The RC technique is further optimized by constraining the number of clusters using the absolute track bias between the ensemble mean (EM) and ensemble spread (ES). Finally, the RC-based deterministic and weighted probabilistic forecasts are compared with the TC track forecasts from traditional methods. It is found that (1) for deterministic TC track forecasts, the RC-based TC track forecasts outperform all other methods at 12–72-h lead times; compared with the skillful EM (118.6 km), the improvements introduced by the use of RC reach up to 10.8% (8.1 km), 10.2% (13.7 km), and 8.7% (20.5 km) at forecast times of 24, 48, and 72 h, respectively. (2) For probabilistic TC track forecasts, RC yields significantly more accurate and discriminative forecasts than traditional equal-weight track forecasts, by increasing the weight of the best cluster, with a decrease of 4.1% in brier score (BS) and an increase of 1.4% in area under the relative operating characteristic curve (AUC). (3) In particular, for cases with recurved tracks, such as typhoons Saudel (2017) and Bavi (2008), RC significantly reduces track errors relative to EM by 56.0% (125.5 km) and 77.7% (192.2 km), respectively. Our results demonstrate that the RC technique not only improves TC track forecasts but also helps to unravel skillful ensemble members, and is likely useful for feature construction in machine learning. [ABSTRACT FROM AUTHOR]

The designation of the cluster number K and the initial centroids is essential for K-modes clustering algorithm. However, most of the improved methods based on K-modes specify the K value manually and generate the initial centroids randomly, which makes the clustering algorithm significantly dependent on human-based decisions and unstable on the iteration time. To overcome this limitation, we propose a cohesive K-modes (CK-modes) algorithm to generate the cluster number K and the initial centroids automatically. Explicitly, we construct a labeled property graph based on index-free adjacency to capture both global and local cohesion of the node in the sample of the input datasets. The cohesive node calculated based on the property similarity is exploited to split the graph to a K-node tree that determines the K value, and then the initial centroids are selected from the split subtrees. Since the property graph construction and the cohesion calculation are only performed once, they account for a small amount of execution time of the clustering operation with multiple iterations, but significantly accelerate the clustering convergence. Experimental validation in both real-world and synthetic datasets shows that the CK-modes algorithm outperforms the state-of-the-art algorithms. [ABSTRACT FROM AUTHOR]

Here, the authors propose a novel two-phase clustering algorithm with a density exploring distance (DED) measure. In the first phase, the fast global K-means clustering algorithm is used to obtain the cluster number and the prototypes. Then, the prototypes of all these clusters and representatives of points belonging to these clusters are regarded as the input data set of the second phase. Afterwards, all the prototypes are clustered according to a DED measure which makes data points locating in the same structure to possess high similarity with each other. In experimental studies, the authors test the proposed algorithm on seven artificial as well as seven UCI data sets. The results demonstrate that the proposed algorithm is flexible to different data distributions and has a stronger ability in clustering data sets with complex non-convex distribution when compared with the comparison algorithms.

In this paper we present a suggested automatic clustering procedure with the main aim to predict the number of clusters of unknown, heterogeneous images. We used the state-of-the-art Fisher-vector for mathematical representation of the images and these vectors were considered as input data points for the clustering algorithm. We implemented a novel variant of K-means, the kernel K-means++, furthermore the min-max kernel K-means plusplus (MMKK++) as clustering method. The proposed approach examines some candidate cluster numbers and uses the law of large numbers in order to choose the optimal cluster size. We conducted experiments on four image sets to demonstrate the efficiency of our solution. The first two image sets are subsets of different popular collections; the third is their union; the fourth is the complete Caltech101 image set.

In this paper, we propose a new approach to recognize the motional patterns of human postures by introducing the data density functional method. Under the framework of the proposed method, sensed time signals will be mapped into specific physical spaces. The most probable cluster number within the specific physical space can be determined according to the principle of energy stability. Then, each corresponding cluster boundary can be measured by searching for the local lowest energy level. Finally, the configuration of the clusters in the space will characterize the most probable states of the motional patterns. The direction of state migration and the corresponding transition region between these states then constitute a significant motional feature in the specific space. Differing from conventional methods, only a single tri-axial gravitational sensor was employed for data acquirement in our hardware scheme. By combining the motional feature and the sensor architecture as prior information, experimental results verified that the most probable states of the motional patterns can be successfully classified into four common human postures of daily life. Furthermore, error motions and noise only offer insignificant influences. Eventually, the proposed approach was applied on a simulation of turning-over situations, and the results show its potential on the issue of elderly and infant turning-over monitoring. [ABSTRACT FROM AUTHOR]

To overcome the main problem of the cluster number in many clustering applications, a new clustering approach with improved morphology similarity distance and the novel cluster validity index is proposed in this paper. An optimized morphology similarity distance based on the Standard Euclidean distance and ReliefF algorithm is used to create a new validity index, which can balance the intra-cluster consistency and inter-cluster consistency. The proposed validity index is combined with fuzzy C-means to produce a creative algorithm simply named the OMS-OSC algorithm. Experimental results obtained using different artificial data sets and real-world data sets show that the new algorithm can not only yield good performance but also detect the correct cluster number. [ABSTRACT FROM AUTHOR]

The role of hydrogen bond interaction on the structural orientation of pure alkanolamines, monoethanolamine (MEA), diethanolamine (DEA), and methyl diethanolamine (MDEA), was under consideration by computer simulation, molecular dynamics simulation as well as density functional theory. Molecular dynamics simulations were carried out at different pressures and temperatures to discover structural stability. A comprehensive force field validation was performed to obtain accurate density. It was also observed that independent of pressure and temperature, MEA molecules prefer head-tail parallel orientation, DEA molecules show both vertical and parallel orientation, and MDEA molecules favor parallel arrangement in liquid state. The probability of hydrogen bond as a substantial intermolecular interaction was also investigated and substantial sites were evaluated. Intermolecular and intramolecular energy, hydrogen bond strength, and transport properties, as well as cluster formation, were investigated to offer an irrefutable proof on structural orientation. Ab initio computation by analyzing frequency, stabilization energy, and electron location function confirmed similar orientation, proposed by molecular dynamics simulation. [ABSTRACT FROM AUTHOR]

In this paper, we propose a new approach to recognize the motional patterns of human postures by introducing the data density functional method. Under the framework of the proposed method, sensed time signals will be mapped into specific physical spaces. The most probable cluster number within the specific physical space can be determined according to the principle of energy stability. Then, each corresponding cluster boundary can be measured by searching for the local lowest energy level. Finally, the configuration of the clusters in the space will characterize the most probable states of the motional patterns. The direction of state migration and the corresponding transition region between these states then constitute a significant motional feature in the specific space. Differing from conventional methods, only a single tri-axial gravitational sensor was employed for data acquirement in our hardware scheme. By combining the motional feature and the sensor architecture as prior information, experimental results verified that the most probable states of the motional patterns can be successfully classified into four common human postures of daily life. Furthermore, error motions and noise only offer insignificant influences. Eventually, the proposed approach was applied on a simulation of turning-over situations, and the results show its potential on the issue of elderly and infant turning-over monitoring.

The Adaptive Neuro-Fuzzy Inference System (ANFIS) modeling technique has shown the ability to model nonlinear system at high accuracy. The initial Fuzzy Inference System (FIS) is required before ANFIS parameters' training. The generation of initial FIS requires the number of inputs, number of clusters and cluster centers to be determined. Instead of determining the above parameters by heuristic approach, the number and variable of inputs of an Electro-Hydraulic Actuator (EHA) system are determined through mathematical derivation of the system. The number of clusters for each input is determined via fuzzy gap statistic, which applies Fuzzy C-means (FCM) instead of k-means of gap statistic in calculating the within cluster dispersion. FCM clustering also provides the center of each cluster identified. Information of the input variables and cluster numbers as well as cluster centers are used to obtain initial FIS. ANFIS hybrid training algorithm is used to train the initial FIS for parameters. The result shows that the model obtained using ANFIS approach can estimate the EHA's system response at high precision. [ABSTRACT FROM AUTHOR]

In cluster analysis, finding the number of clusters, K, for a given dataset is an important yet very tricky task, simply for the facts that there is no universally accepted correct or wrong answer for most real world problems and it all depends on the context and purpose of a cluster study. Numerous methods have been developed for estimating K, but most are not widely used in practice due to their poor performance. Thus, it is still quite common that human user is required to select a specific value or a range for K for many clustering methods before they are used. Inappropriate predetermination for K can result in poor clustering results. This paper presents a new method for estimating the most probable number of clusters automatically. It firstly calculates the length of constant similarity intervals, L, and then considers the longest ones as the representations of the most probable numbers of the clusters under the set context and the chosen similarity measure. An error function is defined to measure and evaluate the goodness of estimations. The proposed method has been tested on 3 synthetic datasets and 8 real-world benchmark datasets, and compared with some other popular methods including particularly the TwoStep implemented in IBM/SPSS Modeler software package. The experimental results showed that the proposed method is able to find the "desired" predominant number of clusters for all the simulated datasets and most of the benchmark datasets, and the statistical tests indicate that our method is significantly better. [ABSTRACT FROM AUTHOR]

Abstract: This paper proposes a new kind of -means algorithms for clustering analysis with three frequency sensitive (data) discrepancy metrics in the cases that the exact number of clusters in a dataset is not pre-known. That is, by setting the number k of seed-points for learning clusters to be larger than the true number of actual clusters in the dataset, i.e., , these algorithms can locate the centers of actual clusters by converged seed-points, respectively, with the extra seed-points corresponding to empty clusters, namely containing no winning points in the competition according to the underlying frequency sensitive discrepancy metrics. It is demonstrated by the experiments on both synthetic and real-world datasets that these three new -means clustering algorithms can detect the number of actual clusters in a dataset with a classification accuracy rate as high as or higher than that of the original -means algorithm. Moreover, they converge more quickly than the original one. [Copyright &y& Elsevier]

Abstract: This paper proposed an improved partition coefficient (IPC) index by turning two disadvantages of partition coefficient (PC) index into advantages. Comparisons between IPC index and PC index and two existed cluster validity indexes are conducted on four real data sets. Experimental results show that IPC is able to identify the cluster number underlying the data set in the case that PC index is unable to do and outperforms the two existed cluster validity indexes. [Copyright &y& Elsevier]

Abstract: This paper introduces k′-means algorithm that performs correct clustering without pre-assigning the exact number of clusters. This is achieved by minimizing a suggested cost-function. The cost-function extends the mean-square-error cost-function of k-means. The algorithm consists of two separate steps. The first is a pre-processing procedure that performs initial clustering and assigns at least one seed point to each cluster. During the second step, the seed-points are adjusted to minimize the cost-function. The algorithm automatically penalizes any possible winning chances for all rival seed-points in subsequent iterations. When the cost-function reaches a global minimum, the correct number of clusters is determined and the remaining seed points are located near the centres of actual clusters. The simulated experiments described in this paper confirm good performance of the proposed algorithm. [Copyright &y& Elsevier]

The existing Rival Penalized Competitive Learning (RPCL) algorithm and its variants have provided an attractive way to perform data clustering without knowing the exact number of clusters. However, their performance is sensitive to the preselection of the rival delearning rate. In this paper, we further investigate the RPCL and present a mechanism to control the strength of rival penalization dynamically. Consequently, we propose the Rival Penalization Controlled Competitive Learning (RPCCL) algorithm and its stochastic version. In each of these algorithms, the selection of the delearning rate is circumvented using a novel technique. We compare the performance of RPCCL to RPCL in Gaussian mixture clustering and color image segmentation, respectively. The experiments have produced the promising results. [ABSTRACT FROM AUTHOR]

Expectation-Maximization (EM) algorithm [10] has been extensively used in density mixture clustering problems, but it is unable to perform model selection automatically. This paper, therefore, proposes to learn the model parameters via maximizing a weighted likelihood. Under a specific weight design, we give out a Rival Penalized Expectation-Maximization (RPEM) algorithm, which makes the components in a density mixture compete each other at each time step. Not only are the associated parameters of the winner updated to adapt to an input, but also all rivals' parameters are penalized with the, strength proportional to the corresponding posterior density probabilities. Compared to the EM algorithm [10], the RPEM is able to fade out the redundant densities from a density mixture during the learning process. Hence, it can automatically select an appropriate number of densities in density mixture clustering. We experimentally demonstrate its outstanding performance on Gaussian mixtures and color image segmentation problem. Moreover, a simplified version of RPEM generalizes our recently proposed RPCCL algorithm [8] so that it is applicable to elliptical clusters as well with any input proportion. Compared to the existing heuristic RPCL [25] and its variants, this generalized RPCCL (G-RPCCL) circumvents the difficult preselection of the so-called delearning rate. Additionally, a special setting of the G-RPCCL not only degenerates to RPCL and its Type A variant, but also gives a guidance to choose an appropriate delearning rate for them. Subsequently, we propose a stochastic version of RPCL and its Type A variant, respectively, in which the difficult selection problem of delearning rate has been novelly circumvented. The experiments show the promising results of this stochastic implementation. [ABSTRACT FROM AUTHOR]

This paper presents a generalized version of the conventional k-means clustering algorithm [Proceedings of 5th Berkeley Symposium on Mathematical Statistics and Probability, 1, University of California Press, Berkeley, 1967, p. 281]. Not only is this new one applicable to ellipse-shaped data clusters without dead-unit problem, but also performs correct clustering without pre-assigning the exact cluster number. We qualitatively analyze its underlying mechanism, and show its outstanding performance through the experiments. [Copyright &y& Elsevier]