Jenni Raitoharju

Evaluating the performance of artificial neural networks for the classification of freshwater benthic macroinvertebrates

Abstract Macroinvertebrates form an important functional component of aquatic ecosystems. Their ability to indicate various types of anthropogenic stressors is widely recognized which has made them an integral component of freshwater biomonitoring. The use of macroinvertebrates in biomonitoring is dependent on manual taxa identification which is currently a time-consuming and cost-intensive process conducted by highly trained taxonomical experts. Automated taxa identification of macroinvertebrates is a relatively recent research development. Previous studies have displayed great potential for solutions to this demanding data mining application. In this research we have a collection of 1350 images from eight different macroinvertebrate taxa and the aim is to examine the suitability of artificial neural networks (ANNs) for automated taxa identification of macroinvertebrates. More specifically, the focus is drawn on different training algorithms of Multi-Layer Perceptron (MLP), probabilistic neural network (PNN) and Radial Basis Function network (RBFN). We performed thorough experimental tests and we tested altogether 13 training algorithms for MLPs. The best classification accuracy of MLPs, 95.3%, was obtained by two conjugate gradient backpropagation variations and scaled conjugate gradient backpropagation. For PNN 92.8% and for RBFN 95.7% accuracies were achieved. The results show how important a proper choice of ANN is in order to obtain high accuracy in the automated taxa identification of macroinvertebrates and the obtained model can outperform the level of identification which is made by a taxonomist.

An evolutionary feature synthesis approach for content-based audio retrieval

A vast amount of audio features have been proposed in the literature to characterize the content of audio signals. In order to overcome specific problems related to the existing features (such as lack of discriminative power), as well as to reduce the need for manual feature selection, in this article, we propose an evolutionary feature synthesis technique with a built-in feature selection scheme. The proposed synthesis process searches for optimal linear/nonlinear operators and feature weights from a pre-defined multi-dimensional search space to generate a highly discriminative set of new (artificial) features. The evolutionary search process is based on a stochastic optimization approach in which a multi-dimensional particle swarm optimization algorithm, along with fractional global best formation and heterogeneous particle behavior techniques, is applied. Unlike many existing feature generation approaches, the dimensionality of the synthesized feature vector is also searched and optimized within a set range in order to better meet the varying requirements set by many practical applications and classifiers. The new features generated by the proposed synthesis approach are compared with typical low-level audio features in several classification and retrieval tasks. The results demonstrate a clear improvement of up to 15–20% in average retrieval performance. Moreover, the proposed synthesis technique surpasses the synthesis performance of evolutionary artificial neural networks, exhibiting a considerable capability to accurately distinguish among different audio classes.

Limited random walk algorithm for big graph data clustering

Graph clustering is an important technique to understand the relationships between the vertices in a big graph. In this paper, we propose a novel random-walk-based graph clustering method. The proposed method restricts the reach of the walking agent using an inflation function and a normalization function. We analyze the behavior of the limited random walk procedure and propose a novel algorithm for both global and local graph clustering problems. Previous random-walk-based algorithms depend on the chosen fitness function to find the clusters around a seed vertex. The proposed algorithm tackles the problem in an entirely different manner. We use the limited random walk procedure to find attractor vertices in a graph and use them as features to cluster the vertices. According to the experimental results on the simulated graph data and the real-world big graph data, the proposed method is superior to the state-of-the-art methods in solving graph clustering problems. Since the proposed method uses the embarrassingly parallel paradigm, it can be efficiently implemented and embedded in any parallel computing environment such as a MapReduce framework. Given enough computing resources, we are capable of clustering graphs with millions of vertices and hundreds millions of edges in a reasonable time.

Training Radial Basis Function Neural Networks for Classification via Class-Specific Clustering

In training radial basis function neural networks (RBFNNs), the locations of Gaussian neurons are commonly determined by clustering. Training inputs can be clustered on a fully unsupervised manner (input clustering), or some supervision can be introduced, for example, by concatenating the input vectors with weighted output vectors (input-output clustering). In this paper, we propose to apply clustering separately for each class (class-specific clustering). The idea has been used in some previous works, but without evaluating the benefits of the approach. We compare the class-specific, input, and input-output clustering approaches in terms of classification performance and computational efficiency when training RBFNNs. To accomplish this objective, we apply three different clustering algorithms and conduct experiments on 25 benchmark data sets. We show that the class-specific approach significantly reduces the overall complexity of the clustering, and our experimental results demonstrate that it can also lead to a significant gain in the classification performance, especially for the networks with a relatively few Gaussian neurons. Among other applied clustering algorithms, we combine, for the first time, a dynamic evolutionary optimization method, multidimensional particle swarm optimization, and the class-specific clustering to optimize the number of cluster centroids and their locations.

Data Enrichment in Fine-Grained Classification of Aquatic Macroinvertebrates

The types and numbers of benthic macroinvertebrates found in a water body reflect water quality. Therefore, macroinvertebrates are routinely monitored as a part of freshwater ecological quality assessment. The collected macroinvertebrate samples are identified by human experts, which is costly and time-consuming. Thus, developing automated identification methods that could partially replace the human effort is important. In our group, we have been working toward this goal and, in this paper, we improve our earlier results on automated macroinvertebrate classification obtained using deep Convolutional Neural Networks (CNNs). We apply simple data enrichment prior to CNN training. By rotations and mirroring, we create new images so as to increase the total size of the image database sixfold. We evaluate the effect of data enrichment on Caffe and MatConvNet CNN implementations. The networks are trained either fully on the macroinvertebrate data or first pretrained using ImageNet pictures and then fine-tuned using the macroinvertebrate data. The results show 3-6% improvement, when the enriched data are used. This is an encouraging result, because it significantly narrows the gap between automated techniques and human experts, while it leaves room for future improvements as even the size of the enriched data, about 60000 images, is small compared to data sizes typically required for efficient training of deep CNNs.

Feature synthesis for image classification and retrieval via one-against-all perceptrons

Most existing content-based image retrieval and classification systems rely on low-level features which are automatically extracted from images. However, often these features lack the discrimination power needed for accurate description of the image content, and hence, they may lead to a poor retrieval or classification performance. We propose a novel technique to improve low-level features which uses parallel one-against-all perceptrons to synthesize new features with a higher discrimination power which in turn leads to improved classification and retrieval results. The proposed method can be applied on any database and low-level features as long as some ground-truth information is available. The main merits of the proposed technique are its simplicity and faster computation compared to existing feature synthesis methods. Extensive simulation results show a significant improvement in the features’ discrimination power.

Particle swarm clustering fitness evaluation with computational centroids

Abstract In this paper, we propose a new way to carry out fitness evaluation in dynamic Particle Swarm Clustering (PSC) with centroid-based encoding. Generally, the PSC fitness function is selected among the clustering validity indices and most of them directly depend on the cluster centroids. In the traditional fitness evaluation approach, the cluster centroids are replaced by the centroids proposed by a particle position. We propose to first compute the centroids of the corresponding clusters and then use these computational centroids in fitness evaluation. The proposed way is called Fitness Evaluation with Computational Centroids (FECC). We conducted an extensive set of comparative evaluations and the results show that FECC leads to a clear improvement in clustering results compared to the traditional fitness evaluation approach with most of the fitness functions considered in this study. The proposed approach was found especially beneficial when underclustering is a problem. Furthermore, we evaluated 31 fitness functions based on 17 clustering validity indices using two PSC methods over a large number of synthetic and real data sets with varying properties. We used three different performance criteria to evaluate the clustering quality and found out that the top three fitness functions are Xu index, WB index, and Dunn variant DU 23 applied using FECC. These fitness functions consistently performed well for both PSC methods, for all data distributions, and according to all performance criteria. In all test cases, they were clearly among the better half of the fitness functions and, in the majority of the cases, they were among the top 4 functions. Further guidance for improved fitness function selection in different situations is provided in the paper.

Tut MUVIS image retrieval system proposal for MSR-Bing challenge 2014

This paper presents our system designed for MSR-Bing Image Retrieval Challenge @ ICME 2014. The core of our system is formed by a text processing module combined with a module performing PCA-assisted perceptron regression with random sub-space selection (P2R2S2). P2R2S2 uses Over-Feat features as a starting point and transforms them into more descriptive features via unsupervised training. The relevance score for each query-image pair is obtained by comparing the transformed features of the query image and the relevant training images. We also use a face bank, duplicate image detection, and optical character recognition to boost our evaluation accuracy. Our system achieves 0.5099 in terms of DCG25 on the development set and 0.5116 on the test set.

Evolutionary feature synthesis for content-based audio retrieval

Although there is a wide variety of low-level audio features for content-based audio indexing and retrieval, they may lack the discrimination power needed for accurate description of the aural content, leading into a poor content-based retrieval performance. Furthermore, manual selection of features among a vast collection may easily lead into sub-optimal solutions. In this paper, we propose an evolutionary feature synthesis technique, which co-exists with a feature selection scheme. The synthesis process seeks for the optimal linear / non-linear operators and feature weights from a pre-defined search space, so as to synthesize a highly discriminative set of new (artificial) features from the set of selected features. The evolutionary search process in the multi-dimensional solution space is based on multi-dimensional particle swarm optimization (MD PSO) algorithm, along with a fractional global best formation (FGBF) technique. Unlike in many existing feature generation approaches found in the literature, the dimension of the synthesized feature vector is also optimized during the process. The synthesized features by the proposed approach are compared with original audio descriptors in an extensive set of retrieval tasks. The experimental results clearly demonstrate a crucial improvement of up to 15-25% in the retrieval performance. Moreover, the proposed synthesis technique surpasses the performance of the artificial neural networks for retrieving accurate audio content.

Benchmark database for fine-grained image classification of benthic macroinvertebrates

Abstract Managing the water quality of freshwaters is a crucial task worldwide. One of the most used methods to biomonitor water quality is to sample benthic macroinvertebrate communities, in particular to examine the presence and proportion of certain species. This paper presents a benchmark database for automatic visual classification methods to evaluate their ability for distinguishing visually similar categories of aquatic macroinvertebrate taxa. We make publicly available a new database, containing 64 types of freshwater macroinvertebrates, ranging in number of images per category from 7 to 577. The database is divided into three datasets, varying in number of categories (64, 29, and 9 categories). Furthermore, in order to accomplish a baseline evaluation performance, we present the classification results of Convolutional Neural Networks (CNNs) that are widely used for deep learning tasks in large databases. Besides CNNs, we experimented with several other well-known classification methods using deep features extracted from the data.