Abstract Synthesising appropriate choreographies from music remains an open problem, due to the need to align musical semantics with subjective human motions. We introduce MDLT, a novel approach that frames the choreography generation problem as a translation task. Our method leverages the AIST++ and PhantomDance data sets to learn to translate sequences of audio into corresponding dance poses. We present two variants of MDLT: MDLT-T based on the Transformer architecture, and MDLT-M based on the Mamba architecture, for strong long-horizon sequence modeling. Trained on these data sets, our method enables a robotic arm to dance, and can generalize to humanoid robots through its architecture-agnostic design. Evaluation metrics, including Average Joint Error and Fréchet Inception Distance, consistently demonstrate that, when given a piece of music, MDLT excels at producing realistic and high-quality choreography. The code will be made available upon acceptance.

Robotic networks are increasingly relied upon to perform complex tasks that require efficient scheduling and task allocation to optimize processing power, resource management, and energy use. The primary goal in these systems is to enhance performance by minimizing completion time, energy consumption, and delays, while maximizing resource utilization and task throughput. Numerous studies have examined different aspects of task allocation and scheduling, from static approaches to dynamic models that adapt to real-time conditions. This article presents a comprehensive survey of the methods and strategies used in robotic network systems, considering not only traditional approaches but also the role of emerging technologies, such as cloud, fog, and edge computing. We categorize the literature from three perspectives: 1) architectures and applications; 2) methods; and 3) parameters. Furthermore, we analyze the limitations of each approach and propose directions for future research, with a particular focus on scalability, real-world applicability, and the integration of these technologies in dynamic environments.

Abstract The widely applied JPEG standard has undergone recent efforts using population-based metaheuristic (PBMH) algorithms to optimise quantisation tables (QTs) for specific images. However, user preferences, like an Android developer’s preference for small-size images, are often overlooked, leading to high-quality images with large file sizes. Another limitation is the lack of comprehensive coverage in current QTs, failing to accommodate all possible combinations of file size and quality. Therefore, this paper aims to propose three distinct contributions. First, to include the user’s opinion in the compression process, the file size of the output image can be controlled by a user in advance. To this end, we propose a novel objective function for population-based JPEG image compression. Second, we suggest a novel representation to tackle the lack of comprehensive coverage. Our proposed representation can not only provide more comprehensive coverage but also find the proper value for the quality factor for a specific image without any background knowledge. Both representation and objective function changes are independent of the search strategies and can be used with any population-based metaheuristic (PBMH) algorithm. Therefore, as the third contribution, we also provide a comprehensive benchmark on 22 state-of-the-art and recently-introduced PBMH algorithms on our new formulation of JPEG image compression. Our extensive experiments on different benchmark images and in terms of different criteria show that our novel formulation for JPEG image compression can work effectively.

Networks found with neural architecture search (NAS) achieve the state-of-the-art performance in a variety of tasks, out-performing human-designed networks. However, most NAS methods heavily rely on human-defined assumptions that constrain the search: architecture's outer skeletons, number of layers, parameter heuristics, and search spaces. In addition, common search spaces consist of repeatable modules (cells) instead of fully exploring the architecture's search space by designing entire architectures (macro-search). Imposing such constraints requires deep human expertise and restricts the search to predefined settings. In this article, we propose less constrained macro-neural architecture search (LCMNAS), a method that pushes NAS to less constrained search spaces by performing macro-search without relying on predefined heuristics or bounded search spaces. LCMNAS introduces three components for the NAS pipeline: 1) a method that leverages information about well-known architectures to autonomously generate complex search spaces based on weighted directed graphs (WDGs) with hidden properties; 2) an evolutionary search strategy that generates complete architectures from scratch; and 3) a mixed-performance estimation approach that combines information about architectures at the initialization stage and lower fidelity estimates to infer their trainability and capacity to model complex functions. We present experiments in 14 different datasets showing that LCMNAS is capable of generating both cell and macro-based architectures with minimal GPU computation and state-of-the-art results. Moreover, we conduct extensive studies on the importance of different NAS components in both cell and macro-based settings. The code for reproducibility is publicly available at https://github.com/VascoLopes/LCMNAS.

Energy consumption is a fundamental concern in mobile application development, bearing substantial significance for both developers and end-users. Main objective of this research is to propose a novel neural network-based framework, enhanced by a metaheuristic approach, to achieve robust energy prediction in the context of mobile app development. The metaheuristic approach here aims to achieve two goals: 1) identifying suitable learning algorithms and their corresponding hyperparameters, and 2) determining the optimal number of layers and neurons within each layer. Moreover, due to limitations in accessing certain aspects of a mobile phone, there might be missing data in the data set, and the proposed framework can handle this. In addition, we conducted an optimal algorithm selection strategy, employing 13 base and advanced metaheuristic algorithms, to identify the best algorithm based on accuracy and resistance to missing values. The representation in our proposed metaheuristic algorithm is variable-size, meaning that the length of the candidate solutions changes over time. We compared the algorithms based on the architecture found by each algorithm at different levels of missing values, accuracy, F-measure, and stability analysis. Additionally, we conducted a Wilcoxon signed-rank test for statistical comparison of the results. The extensive experiments show that our proposed approach significantly improves energy consumption prediction. Particularly, the JADE algorithm, a variant of differential evolution (DE), DE, and the covariance matrix adaptation evolution strategy deliver superior results under various conditions and across different missing value levels.

One important application of machine learning is the analysis of electroencephalographic recordings (EEG). Such oscillatory signals are noisy, non-stationary, full of artifacts, contain transients, and chaotic transitions between meta-stable states. EEG data rarely consists of recordings obtained for more than 100 patients with some brain disorder. This makes objective, reliable clinical diagnosis of many mental disorders, such as schizophrenia, very difficult. The standard approach splits the frequency of EEG oscillations focusing on five classical bands: delta, theta, alpha, beta, and gamma. In this paper, we investigate three main questions: (i) Are there certain frequencies that will allow for a reliable diagnosis of schizophrenia, and (ii) if they are, is there a connection to what is known about their neural basis? (iii) How long segments of EEG recordings are sufficient for reliable classification? We filter EEG signals, obtaining a set of 64 very narrow 1Hz frequency bands. A genetic algorithm, with a simple k-NN classifier, finds an optimal combination of these bands. The method is sufficiently simple to be used in clinical settings. A public schizophrenia EEG data set, containing 60 seconds of EEG recordings in the resting state, with only 16 electrodes, for 45 adolescent patients and 39 healthy controls, was used for testing. Signal duration of about 30 seconds enables to reach an accuracy of over 96% in 5-fold cross-validation tests. We compare our results to much more sophisticated state-of-the- art methods and discuss insights gained by our analysis into the brain basis of schizophrenia.

Neural Architecture Search (NAS) methods have been successfully applied to image tasks with excellent results. However, NAS methods are often complex and tend to converge to local minima as soon as generated architectures yield good results. This paper proposes GEA, a novel approach for guided NAS. GEA guides the evolution by exploring the search space by generating and evaluating several architectures in each generation at initialisation stage using a zero-proxy estimator, where only the highest-scoring architecture is trained and kept for the next generation. Subsequently, GEA continuously extracts knowledge about the search space without increased complexity by generating several off-springs from an existing architecture at each generation. Moreover, GEA forces exploitation of the most performant architectures by descendant generation while simultaneously driving exploration through parent mutation and favouring younger architectures to the detriment of older ones. Experimental results demonstrate the effectiveness of the proposed method, and extensive ablation studies evaluate the importance of different parameters. Results show that GEA achieves competitive results on all data sets of NAS-Bench-101, NAS-Bench-201 and TransNAS-Bench-101 benchmarks, as well as in the DARTS search space.

Deploying reinforcement learning agents in the real world can be challenging due to the risks associated with learning through trial and error.We propose a task-agnostic method that leverages small sets of safe and unsafe demonstrations to improve the safety of RL agents during learning.The method compares the current trajectory of the agent with both sets of demonstrations at every step, and filters the trajectory if it resembles the unsafe demonstrations.We perform ablation studies on different filtering strategies and investigate the impact of the number of demonstrations on performance.Our method is compatible with any stand-alone RL algorithm and can be applied to any task.We evaluate our method on three tasks from OpenAI Gym's Mujoco benchmark and two state-of-the-art RL algorithms.The results demonstrate that our method significantly reduces the crash rate of the agent while converging to, and in most cases even improving, the performance of the stand-alone agent.

Sequence models in reinforcement learning require task knowledge to estimate the task policy. This paper presents the hierarchical decision transformer (HDT). HDT is a hierarchical behavior cloning algorithm that improves the performance of transformer methods in imitation learning, improving their robustness to tasks with longer episodes and/or sparse rewards, without requiring task knowledge or user interaction currently present in the state-of-the-art. The high-level mechanism guides the low-level controller through the task by selecting sub-goals for the latter to reach. This sequence replaces the returns-to-go of previous methods, improving its performance overall, especially in tasks with longer episodes and scarcer rewards. We validate our method in multiple tasks of OpenAI Gym, D4RL, and RoboMimic benchmarks. Our method outperforms the baselines in twenty three out of thirty one settings of varied horizons and reward frequencies without prior task knowledge, showing the advantages of the hierarchical model approach for learning from demonstrations using a sequence model. We also evaluate the method on a reaching task on a physical robot.

Neural Architecture Search (NAS) benchmarks significantly improved the capability of developing and comparing NAS methods while at the same time drastically reduced the computational overhead by providing meta-information of trained neural networks. However, tabular benchmarks have several drawbacks that can hinder fair comparisons and provide unreliable results. These usually focus on a small pool of operations in heavily constrained search spaces – usually cell-based neural networks with pre-defined outer-skeletons. In this work, we conducted an empirical analysis of the widely used NAS-Bench-101, NAS-Bench-201 and TransNAS-Bench-101 benchmarks in terms of their generability and how different operations influence the performance of the generated architectures. We found that only a subset of the operation pool is required to generate architectures close to the upper-bound of the performance range. More, the performance distribution is negatively skewed, with many architectures clustered near the upper accuracy bound. Further experiments revealed that convolution layers have the highest impact on the architecture's performance and that specific combinations of operations favor top-scoring architectures. Overall, our results demonstrate the need for benchmarks with greater operation diversity and less constrained search spaces. We provide suggestions for improving future benchmark design and evaluation of NAS methods when using existing benchmarks. The code used to conduct the evaluations is available at https://github.com/VascoLopes/NAS-Benchmark-Evaluation.

Abstract Object pose estimation has multiple important applications, such as robotic grasping and augmented reality. We present a new method to estimate the 6D pose of objects that improves upon the accuracy of current proposals and can still be used in real-time. Our method uses RGB-D data as input to segment objects and estimate their pose. It uses a neural network with multiple heads to identify the objects in the scene, generate the appropriate masks and estimate the values of the translation vectors and the quaternion that represents the objects’ rotation. These heads leverage a pyramid architecture used during feature extraction and feature fusion. We conduct an empirical evaluation using the two most common datasets in the area, and compare against state-of-the-art approaches, illustrating the capabilities of MPF6D. Our method can be used in real-time with its low inference time and high accuracy.

For a robot to perform a task, several algorithms must be executed, sometimes simultaneously. The algorithms can be executed either on the robot itself or, if desired, on a cloud infrastructure. The term cloud infrastructure refers to hardware, storage, abstracted resources, and network resources associated with cloud computing. Depending on the decision of where the algorithms are executed, the overall execution time and memory required for the robot, change accordingly. The price of a robot depends on its storage capacity and computational power, among other factors. We answer the question of how to maintain a given performance and deploy a cheaper robot (lower resources) by allocating computational tasks to the cloud infrastructure depending on memory, computational power, and communication constraints. Even for a fixed robot, our model provides a way to achieve optimal overall performance. We provide a general model for optimal algorithm allocation decision under certain constraints. We illustrate the model with simulation results. The main advantage of our model is that it provides optimal task allocation simultaneously for memory and time.

Robotic networks can be used to accomplish tasks that exceed the capacity of a single robot. In a robotic network, robots can work together to accomplish a common task. Cloud robotics allows robots to benefit from the massive storage and computing power of the cloud. Previous studies mainly focus on minimizing the cost of resource retrieval by robots by knowing the resource allocation in advance. Duplicating algorithms on multiple nodes can reduce the total time required to execute a task. We address the question of which algorithms should be duplicated and where the duplicates should be placed to improve overall performance. We have developed a procedure to answer wherein a robotic network cloud system should algorithms be executed and whether they should be duplicated to achieve optimal performance in terms of overall task execution time for all robots. Our proposed duplication procedure is optimal in the sense that the number of duplicated algorithms is minimal, while the result provides minimal overall completion time for all robots.

Blockchain is a disruptive technology, normally used within financial applications, however, it can be very beneficial in certain robotic contexts, such as when an immutable register of events is required. Among the several properties of Blockchain that can be useful within robotic environments, we find not just immutability but also data decentralization, irreversibility, accessibility and non-repudiation. In this paper, we propose an architecture that uses blockchain as a ledger, and smart-contracts for robotic control by using oracles to process data. We show how to register events in a secure way, how it is possible to use smart-contracts to control robots and how to interface with external algorithms for image analysis. The proposed architecture is modular and can be used in multiple contexts such as in manufacturing, network control, robot control, and others, since it is easy to integrate, adapt, maintain and extend to new domains, only requiring new tailored smart-contracts.

Sentiment analysis is a research topic focused on analysing data to extract information related to the sentiment that it causes. Applications of sentiment analysis are wide, ranging from recommendation systems, and marketing to customer satisfaction. Recent approaches evaluate textual content using Machine Learning techniques that are trained over large corpora. However, as social media grown, other data types emerged in large quantities, such as images. Sentiment analysis in images has shown to be a valuable complement to textual data since it enables the inference of the underlying message polarity by creating context and connections. Multimodal sentiment analysis approaches intend to leverage information of both textual and image content to perform an evaluation. Despite recent advances, current solutions still flounder in combining both image and textual information to classify social media data, mainly due to subjectivity, inter-class homogeneity and fusion data differences. In this paper, we propose a method that combines both textual and image individual sentiment analysis into a final fused classification based on AutoML, that performs a random search to find the best model. Our method achieved state-of-the-art performance in the B-T4SA dataset, with 95.19% accuracy.

MaskedFusion is a framework to estimate the 6D pose of objects using RGB-D data, with an architecture that leverages multiple sub-tasks in a pipeline to achieve accurate 6D poses. 6D pose estimation is an open challenge due to complex world objects and many possible problems when capturing data from the real world, e.g., occlusions, truncations, and noise in the data. Achieving accurate 6D poses will improve results in other open problems like robot grasping or positioning objects in augmented reality. MaskedFusion improves the state-of-the-art by using object masks to eliminate non-relevant data. With the inclusion of the masks on the neural network that estimates the 6D pose of an object we also have features that represent the object shape. MaskedFusion is a modular pipeline where each sub-task can have different methods that achieve the objective. MaskedFusion achieved 97.3% on average using the ADD metric on the LineMOD dataset and 93.3% using the ADD-S AUC metric on YCB-Video Dataset, which is an improvement, compared to the state-of-the-art methods.

Monocular 3D reconstruction is a challenging computer vision task that becomes even more stimulating when we aim at real-time performance. One way to obtain 3D reconstruction maps is through the use of Simultaneous Localization and Mapping (SLAM), a recurrent engineering problem, mainly in the area of robotics. It consists of building and updating a consistent map of the unknown environment and, simultaneously, saving the pose of the robot, or the camera, at every given time instant. A variety of algorithms has been proposed to address this problem, namely the Large Scale Direct Monocular SLAM (LSD-SLAM), ORB-SLAM, Direct Sparse Odometry (DSO) or Parallel Tracking and Mapping (PTAM), among others. However, despite the fact that these algorithms provide good results, they are computationally intensive. Hence, in this paper, we propose a modified version of DSO SLAM, which implements code parallelization techniques using OpenMP, an API for introducing parallelism in C, C++ and Fortran programs, that supports multi-platform shared memory multi-processing programming. With this approach we propose multiple directive-based code modifications, in order to make the SLAM algorithm execute considerably faster. The performance of the proposed solution was evaluated on standard datasets and provides speedups above 40% without significant extra parallel programming effort.

Over the last years, deep learning architectures have gained attention by winning important international detection and classification challenges. However, due to high levels of energy consumption, the need to use low-power devices at acceptable throughput performance is higher than ever. This paper tries to solve this problem by introducing energy efficient deep learning based on local training and using low-power mobile GPU parallel architectures, all conveniently supported by the same high-level description of the deep network. Also, it proposes to discover the maximum dimensions that a particular type of deep learning architecture - the stacked autoencoder - can support by finding the hardware limitations of a representative group of mobile GPUs and platforms.

One of the major problems found when developing a 3D recognition system involves the choice of keypoint detector and descriptor. To help solve this problem, we present a new method for the detection of 3D keypoints on point clouds and we perform benchmarking between each pair of 3D keypoint detector and 3D descriptor to evaluate their performance on object and category recognition. These evaluations are done in a public database of real 3D objects. Our keypoint detector is inspired by the behavior and neural architecture of the primate visual system. The 3D keypoints are extracted based on a bottom-up 3D saliency map, that is, a map that encodes the saliency of objects in the visual environment. The saliency map is determined by computing conspicuity maps (a combination across different modalities) of the orientation, intensity, and color information in a bottom-up and in a purely stimulus-driven manner. These three conspicuity maps are fused into a 3D saliency map and, finally, the focus of attention (or keypoint location) is sequentially directed to the most salient points in this map. Inhibiting this location automatically allows the system to attend to the next most salient location. The main conclusions are: with a similar average number of keypoints, our 3D keypoint detector outperforms the other eight 3D keypoint detectors evaluated by achieving the best result in 32 of the evaluated metrics in the category and object recognition experiments, when the second best detector only obtained the best result in eight of these metrics. The unique drawback is the computational time, since biologically inspired 3D keypoint based on bottom-up saliency is slower than the other detectors. Given that there are big differences in terms of recognition performance, size and time requirements, the selection of the keypoint detector and descriptor has to be matched to the desired task and we give some directions to facilitate this choice.

In this paper, we propose a robust detection and tracking method for 3D objects by using keypoint information in a particle filter. Our method consists of three distinct steps: Segmentation, Tracking Initialization and Tracking. The segmentation is made in order to remove all the background information, in order to reduce the number of points for further processing. In the initialization, we use a keypoint detector with biological inspiration. The information of the object that we want to follow is given by the extracted keypoints. The particle filter does the tracking of the keypoints, so with that we can predict where the keypoints will be in the next frame. In a recognition system, one of the problems is the computational cost of keypoint detectors with this we intend to solve this problem. The experiments with PFBIK-Tracking method are done indoors in an office/home environment, where personal robots are expected to operate. The Tracking Error evaluate the stability of the general tracking method. We also quantitatively evaluate this method using a “Tracking Error”. Our evaluation is done by the computation of the keypoint and particle centroid. Comparing our system with the tracking method which exists in the Point Cloud Library, we archive better results, with a much smaller number of points and computational time. Our method is faster and more robust to occlusion when compared to the OpenniTracker.

Deep architectures have been used in transfer learning applications, with the aim of improving the performance of networks designed for a given problem by reusing knowledge from another problem. In this work we addressed the transfer of knowledge between deep networks used as classifiers of digit and shape images, considering cases where only the set of class labels, or only the data distribution, changed from source to target problem. Our main goal was to study how the performance of knowledge transfer between such problems would be affected by varying the number of layers being retrained and the amount of data used in that retraining. Generally, reusing networks trained for a different label set led to better results than reusing networks trained for a different data distribution. In particular, reusing for less classes a network trained for more classes was beneficial for virtually any amount of training data. In all cases, retraining only one layer to save time consistently led to poorer performance. The results obtained when retraining for upright digits a network trained for rotated digits raise the hypothesis that transfer learning could be used to better deal with image classification problems in which only a small amount of labelled data is available for training.

Transfer learning is a process that allows reusing a learning machine trained on a problem to solve a new problem. Transfer learning studies on shallow architectures show low performance as they are generally based on hand-crafted features obtained from experts. It is therefore interesting to study transference on deep architectures, known to directly extract the features from the input data. A Stacked Denoising Autoencoder (SDA) is a deep model able to represent the hierarchical features needed for solving classification problems. In this paper we study the performance of SDAs trained on one problem and reused to solve a different problem not only with different distribution but also with a different tasks. We propose two different approaches: 1) unsupervised feature transference, and 2) supervised feature transference using deep transfer learning. We show that SDAs using the unsupervised feature transference outperform randomly initialized machines on a new problem. We achieved 7% relative improvement on average error rate and 41% on average computation time to classify typed uppercase letters. In the case of supervised feature transference, we achieved 5.7% relative improvement in the average error rate, by reusing the first and second hidden layer, and 8.5% relative improvement for the average error rate and 54% speed up w.r.t the baseline by reusing all three hidden layers for the same data. We also explore transfer learning between geometrical shapes and canonical shapes, we achieved 7.4% relative improvement on average error rate in case of supervised feature transference approach.

This paper announces and discusses the experimental results from the Noisy Iris Challenge Evaluation (NICE), an iris biometric evaluation initiative that received worldwide participation and whose main innovation is the use of heavily degraded data acquired in the visible wavelength and uncontrolled setups, with subjects moving and at widely varying distances. The NICE contest included two separate phases: 1) the NICE.I evaluated iris segmentation and noise detection techniques and 2) the NICE:II evaluated encoding and matching strategies for biometric signatures. Further, we give the performance values observed when fusing recognition methods at the score level, which was observed to outperform any isolated recognition strategy. These results provide an objective estimate of the potential of such recognition systems and should be regarded as reference values for further improvements of this technology, which-if successful-may significantly broaden the applicability of iris biometric systems to domains where the subjects cannot be expected to cooperate.

This letter focuses on the issue of whether risk functionals derived from information-theoretic principles, such as Shannon or Rényi's entropies, are able to cope with the data classification problem in both the sense of attaining the risk functional minimum and implying the minimum probability of error allowed by the family of functions implemented by the classifier, here denoted by min Pe. The analysis of this so-called minimization of error entropy (MEE) principle is carried out in a single perceptron with continuous activation functions, yielding continuous error distributions. In spite of the fact that the analysis is restricted to single perceptrons, it reveals a large spectrum of behaviors that MEE can be expected to exhibit in both theory and practice. In what concerns the theoretical MEE, our study clarifies the role of the parameters controlling the perceptron activation function (of the squashing type) in often reaching the minimum probability of error. Our study also clarifies the role of the kernel density estimator of the error density in achieving the minimum probability of error in practice.

The iris is regarded as one of the most useful traits for biometric recognition and the dissemination of nationwide iris-based recognition systems is imminent. However, currently deployed systems rely on heavy imaging constraints to capture near infrared images with enough quality. Also, all of the publicly available iris image databases contain data correspondent to such imaging constraints and therefore are exclusively suitable to evaluate methods thought to operate on these type of environments. The main purpose of this paper is to announce the availability of the UBIRIS.v2 database, a multisession iris images database which singularly contains data captured in the visible wavelength, at-a-distance (between four and eight meters) and on on-the-move. This database is freely available for researchers concerned about visible wavelength iris recognition and will be useful in accessing the feasibility and specifying the constraints of this type of biometric recognition.

This paper presents an extensive evaluation of reservoir computing for the case of classification problems that do not depend on time.We discuss how it is possible to adapt the reservoir approach to learning for the case of static classification problems. Then we present a set of experiments against K-PLS, MLP with entropic cost function and LS-SVM showing that this approach is quite competitive and has the advantage of having only one parameter to be chosen.

Textual data holds a number of properties that can be taken into account in order to improve compression. Pre-processing deals with these properties by applying a number of transformations that make the redundancy "more visible" to the compressor. One of the most commonly used concepts in text pre-processing is called capital conversion. Words with capital letters are converted to their lowercase versions while signaling the change with a flag. This way not only context similarities are increased but also dictionaries used for word replacement only need to contain words in their lowercase versions. Word replacement consists of replacing words with shorter codes which are references to their location in a dictionary.

Hierarchical clustering is a stepwise clustering method usually based on proximity measures between objects or sets of objects from a given data set. The most common proximity measures are distance measures. The derived proximity matrices can be used to build graphs, which provide the basic structure for some clustering methods. We present here a new proximity matrix based on an entropic measure and also a clustering algorithm (LEGClust) that builds layers of subgraphs based on this matrix, and uses them and a hierarchical agglomerative clustering technique to form the clusters. Our approach capitalizes on both a graph structure and a hierarchical construction. Moreover, by using entropy as a proximity measure we are able, with no assumption about the cluster shapes, to capture the local structure of the data, forcing the clustering method to reflect this structure. We present several experiments on artificial and real data sets that provide evidence on the superior performance of this new algorithm when compared with competing ones.

This paper focuses on noncooperative iris recognition, i.e., the capture of iris images at large distances, under less controlled lighting conditions, and without active participation of the subjects. This increases the probability of capturing very heterogeneous images (regarding focus, contrast, or brightness) and with several noise factors (iris obstructions and reflections). Current iris recognition systems are unable to deal with noisy data and substantially increase their error rates, especially the false rejections, in these conditions. We propose an iris classification method that divides the segmented and normalized iris image into six regions, makes an independent feature extraction and comparison for each region, and combines each of the dissimilarity values through a classification rule. Experiments show a substantial decrease, higher than 40 percent, of the false rejection rates in the recognition of noisy iris images.

In this paper we analyze the relationship between the accuracy of the segmentation algorithm and the error rates of typical iris recognition systems. We selected 1000 images from the UBIRIS database that the segmentation algorithm can accurately segment and artificially introduced segmentation inaccuracies. We repeated the recognition tests and concluded about the strong relationship between the errors in the pupil segmentation and the overall false reject rate. Based on this fact, we propose a method to identify these inaccuracies.

Entropy-based cost functions are enjoying a growing attractiveness in unsupervised and supervised classification tasks. Better performances in terms both of error rate and speed of convergence have been reported. In this letter, we study the principle of error entropy minimization (EEM) from a theoretical point of view. We use Shannon's entropy and study univariate data splitting in two-class problems. In this setting, the error variable is a discrete random variable, leading to a not too complicated mathematical analysis of the error entropy. We start by showing that for uniformly distributed data, there is equivalence between the EEM split and the optimal classifier. In a more general setting, we prove the necessary conditions for this equivalence and show the existence of class configurations where the optimal classifier corresponds to maximum error entropy. The presented theoretical results provide practical guidelines that are illustrated with a set of experiments with both real and simulated data sets, where the effectiveness of EEM is compared with the usual mean square error minimization.

In a classification problem, improved accuracy can be obtained in many situations by using the combination of several classifiers instead of a single one. Turner and Gosh (1999) derived the error reduction that can be obtained by combining unbiased classifiers with independent errors using a simple average. We present an extension of this result by finding the improvement obtained when combining classifiers using weighted average. We also prove that for unbiased classifiers with independent errors the best combination of N classifiers corresponds to a weighted average, where the combination coefficient of each classifier is equal to 1/N. This means that in these cases the simple average should be used. We present experiments illustrating our results.