Reliable recognition and localization of surgical instruments in endoscopic video recordings are foundational for a wide range of applications in computer- and robot-assisted minimally invasive surgery (RAMIS), including surgical training, skill assessment, and autonomous assistance. However, robust performance under real-world conditions remains a significant challenge. Incorporating surgical context - such as the current procedural phase - has emerged as a promising strategy to improve robustness and interpretability. To address these challenges, we organized the Surgical Procedure Phase, Keypoint, and Instrument Recognition (PhaKIR) sub-challenge as part of the Endoscopic Vision (EndoVis) challenge at MICCAI 2024. We introduced a novel, multi-center dataset comprising thirteen full-length laparoscopic cholecystectomy videos collected from three distinct medical institutions, with unified annotations for three interrelated tasks: surgical phase recognition, instrument keypoint estimation, and instrument instance segmentation. Unlike existing datasets, ours enables joint investigation of instrument localization and procedural context within the same data while supporting the integration of temporal information across entire procedures. We report results and findings in accordance with the BIAS guidelines for biomedical image analysis challenges. The PhaKIR sub-challenge advances the field by providing a unique benchmark for developing temporally aware, context-driven methods in RAMIS and offers a high-quality resource to support future research in surgical scene understanding.

The application of large language models in health-care presents unique challenges, particularly in non-English contexts where linguistic and cultural nuances significantly impact model effectiveness. In this work, we introduce a novel benchmark for evaluating medical language models in Brazilian Portuguese, addressing a critical gap in AI assessment for healthcare applications. This benchmark is built upon Brazilian medical aptitude tests spanning 2011–2024, enabling extensive evaluation of both specialist and general large language models. Our findings demonstrate that despite advancements in language model capabilities, significant gaps remain in their ability to reason effectively about medical knowledge in Brazilian Portuguese. This benchmark establishes a proper foundation for evaluating and advancing medical language models in Portuguese, creating a standardized framework to guide development toward more effective, equitable, and culturally appropriate AI systems for healthcare in Brazil.

This paper explores the potential of quantum ma-chine learning for breast cancer detection. We designed a binary classification approach using the BreastMNIST dataset and segmented mass regions derived from the BCDR dataset. A quanvolutional layer is employed as a quantum feature extractor, interfaced with elements of classical neural networks, to enhance the detection of malignant and benign patterns in breast tissue. The hybrid quanvolutional neural network aims to mitigate challenges associated with traditional machine learning models, such as feature sparsity and data imbalance. This architecture employs a simple yet efficient design that integrates the strengths of both quantum computing and classical methods, reducing computational complexity while maintaining performance. Re-sults demonstrate the potential of quanvolutions in diagnostic accuracy, offering a promising framework for integrating quan-tum computing in medical imaging. This approach provides an optimized solution that balances quantum processing with classical systems for more effective and scalable applications.

Image super-resolution (SR) focuses on reconstructing high-resolution images from their low-resolution counter-parts, often affected by sensor limitations or environmental factors. Convolutional Neural Networks (CNNs) are state-of-the-art for SR tasks but computationally heavy. This paper introduces a novel CRMN (Convolutional Recurrent Mixer Network), a hybrid deep learning-based SR technique designed to address the complexity of CNNs, which is validated in the context of meteorological radar images. Experiments on public benchmark datasets (Berkley432 and T291) and our newly manually collected precipitation dataset from the Meteorological Research Institute (IPMET) show that our CRMN model provides competitive results compared to leading SR methods with significantly fewer parameters, making it a promising and practical solution for SR applications, particularly radar meteorology.

Quantum optimization algorithms, particularly the Quantum Approximate Optimization Algorithm (QAOA), have significantly addressed combinatorial optimization problems. While QAOA has been applied to various NP-hard problems, such as Max-Cut and the Traveling Salesman Problem (TSP), its application to the Degree-Constrained Minimum Spanning Tree (DCMST) problem remains unexplored. Inspired by Fowler’s formulation, this work presents the first implementation of the DCMST Hamiltonian within QAOA and insights from its benefits to graph-based machine learning algorithms. We investigated two approaches: (i) the standard QAOA ansatz with an X mixer and (ii) a warm-started QAOA utilizing classical preprocessing to enhance convergence. We used these strategies to provide numerical results for instances with 3 and 4 nodes, employing the COBYLA optimizer and metaheuristic optimization techniques. Our findings serve as a proof of concept, demonstrating the feasibility of applying QAOA to this problem; however, the number of qubits scales as O(N2) with the number of nodes, limiting scalability. Current research focuses on developing more efficient implementations that encode the same number of binary variables using fewer qubits. This study contributes to the expanding field of quantum optimization, highlighting the potential of hybrid quantum-classical algorithms, especially with resource-efficient mixers and warm-starting techniques, to solve complex combinatorial problems and advance the development of scalable quantum algorithms.

Thermal imaging of venous leg ulcers has helped clinicians make informed wound management decisions. However, thermal cameras are not available in most clinics. To overcome this, we propose a pilot test using deep learning to estimate thermal images from RGB data of the ulcers. Our approach employs stable diffusion techniques, e.g., DreamBooth, LoRA, and ControlNet, to create thermal images from RGB data, addressing the limitations of cost and accessibility in conventional thermal imaging to assist clinicians in assessing the ulcers. While the images’ visualization appears helpful, achieving an average structural similarity index measure (SSIM) score of 0.84, this study has yet to test their suitability for a computerized assessment of chronic wounds.

This work proposes a chatbot architecture for screening Parkinson’s disease (PD) using the Telegram platform. By leveraging large language models (LLMs) as agents, our chatbot can run the voice based screening inspections for computer-assisted PD diagnosis. The chatbot guides users through vocal tests that analyze phonemes, a method based on the potential of vocal biomarkers to indicate PD. This innovation makes PD screening more accessible and user-friendly, particularly in underserved regions, offering a cost-effective tool for early intervention and improved patient care.

Recent advances in generative capabilities provided by large language models have reshaped technology research and human society’s cognitive abilities, bringing new innovative capacities to artificial intelligence solutions. However, the size of such models has raised several concerns regarding their alignment with hardware-limited resources. This paper presents a comprehensive study on training Portuguese-focused Small Language Models (SLMs). We have developed a unique dataset for training our models and employed full fine-tuning, as well as PEFT approaches for comparative analysis. We used Microsoft’s Phi and Google’s Gemma as base models to create our own, named PhiBode and GemBode. These models range from approximately 1 billion to 7 billion parameters, with a total of ten models developed. Our findings provide valuable insights into the performance and applicability of these models, contributing significantly to the field of Portuguese language processing. This research is a step forward in understanding and improving the performance of SLMs in Portuguese. The comparative analysis of the models provides a clear benchmark for future research in this area. The results demonstrate the effectiveness of our training methods and the potential of our models for various applications. This paper significantly contributes to language model training, particularly for the Portuguese language.

Abstract Recent advancements in deep learning generative models have raised concerns as they can create highly convincing counterfeit images and videos. This poses a threat to people's integrity and can lead to social instability. To address this issue, there is a pressing need to develop new computational models that can efficiently detect forged content and alert users to potential image and video manipulations. This paper presents a comprehensive review of recent studies for deepfake content detection using deep learning‐based approaches. We aim to broaden the state‐of‐the‐art research by systematically reviewing the different categories of fake content detection. Furthermore, we report the advantages and drawbacks of the examined works, and prescribe several future directions towards the issues and shortcomings still unsolved on deepfake detection.

Abstract Regularization helps to improve machine learning techniques by penalizing the models during training. Such approaches act in either the input, internal, or output layers. Regarding the latter, label smoothing is widely used to introduce noise in the label vector, making learning more challenging. This work proposes a new label regularization method, Random Label Smoothing, that attributes random values to the labels while preserving their semantics during training. The idea is to change the entire label into fixed arbitrary values. Results show improvements in image classification and super-resolution tasks, outperforming state-of-the-art techniques for such purposes.

Despite the recent success of machine learning algorithms, most models face drawbacks when considering more complex tasks requiring interaction between different sources, such as multimodal input data and logical time sequences. On the other hand, the biological brain is highly sharpened in this sense, empowered to automatically manage and integrate such streams of information. In this context, this work draws inspiration from recent discoveries in brain cortical circuits to propose a more biologically plausible self-supervised machine learning approach. This combines multimodal information using intra-layer modulations together with Canonical Correlation Analysis, and a memory mechanism to keep track of temporal data, the overall approach termed Canonical Cortical Graph Neural networks. This is shown to outperform recent state-of-the-art models in terms of clean audio reconstruction and energy efficiency for a benchmark audio-visual speech dataset. The enhanced performance is demonstrated through a reduced and smother neuron firing rate distribution. suggesting that the proposed model is amenable for speech enhancement in future audio-visual hearing aid devices.

Remote sensing change detection aims to identify significant variations in aerial image acquisition during different time frames. A decisive change detector is necessary to filter out the interest regions, such as recent urban buildings and changed vegetation, from undesired detections, i.e., artifacts generated by misregistration and illumination changes. To overcome common change detection problems (false positive and false negative alarms) and also processing overhead, this manuscript proposes a lighTweIght Temporal Attention Network, aka TITAN, which comprises a partial-siamese deep learning-based change detector that leverages the natural capacity of an encoder-decoder framework to extract different levels of feature information from its input data. To assist the process of combining the meaningful encoded spatial-temporal information with its corresponding semantic decoded counterpart, we also propose the Temporal Change Attention Module (TCAM). Although TCAM does not explicitly account for non-local spatial changes, results support the claim that it implicitly helped TITAN in the matter. The experimental results show the proposed approach overcomes three out of four state-of-the-art techniques in terms of overall average F-measure, Intersection over Union, and Percentage of Wrong Classification measures calculated over SZATAKI, Onera, LEVIR, and SYSU-CD remote sensing change detection datasets, with the lowest overhead.

Deep learning architectures have achieved promising results in different areas (e.g., medicine, agriculture, and security). However, using those powerful techniques in many real applications becomes challenging due to the large labeled collections required during training. Several works have pursued solutions to overcome it by proposing strategies that can learn more for less, e.g., weakly and semi-supervised learning approaches. As these approaches do not usually address memorization and sensitivity to adversarial examples, this paper presents three deep metric learning approaches combined with Mixup for incomplete-supervision scenarios. We show that some state-of-the-art approaches in metric learning might not work well in such scenarios. Moreover, the proposed approaches outperform most of them in different datasets.

Identifying anomalies has become one of the primary strategies towards security and protection procedures in computer networks. In this context, machine learning-based methods emerge as an elegant solution to identify such scenarios and learn irrelevant information so that a reduction in the identification time and possible gain in accuracy can be obtained. This paper proposes a novel feature selection approach called Finite Element Machines for Feature Selection (FEMa-FS), which uses the framework of finite elements to identify the most relevant information from a given dataset. Although FEMa-FS can be applied to any application domain, it has been evaluated in the context of anomaly detection in computer networks. The outcomes over two datasets showed promising results.

Machine learning techniques have achieved promising results in detecting attacks in computer networks, particularly ensemble learning methods, improving individual classifier’s performance. This work focuses on building an ensemble of classifiers to minimize the computational cost to some extent. A diversity-driven pruning method was applied to create stackings using a combination of k-Nearest Neighbors, Decision Trees, Support Vector Machines, and Neural Networks, and validated on six differents datasets. An average accuracy of 99.94% and a reduction in the processing time of 97.34% are reported with heterogeneous ensembles, highlighting the robustness of the proposed approach.

Abstract Biometric recognition provides straightforward methods to deal with the problem of identifying people under certain circumstances. Additionally, a well‐calibrated biometric system enhances security policies and prevents malicious attempts, such as fraud or identity theft. Deep learning has arisen to foster the problem by extracting high‐level features that compose the so‐called ‘user fingerprint’, that is, digital identification of a particular individual. Nevertheless, personal identification is not a trivial task, as many traits might define an individual, varying according to the task's domain. An exciting way to overcome such a problem is to employ handwritten dynamics, which are hand‐ and motor‐based signals from an individual's writing style and obtained through a biometric smartpen. In this work, we propose using such signals to identify an individual through convolutional neural networks. Essentially, the proposed work uses a neighbour‐based bag‐of‐samplings procedure to sample the signals to a fixed size and feeds them into a neural network responsible for extracting their features and further classifying them. The experiments were conducted over two handwritten dynamic datasets, NewHandPD and SignRec, and established new fruitful state‐of‐the‐art concerning these particular datasets and the corresponding context.

Real-time detection of possible deforestation of urban landscapes is an essential task for many urban forest monitoring services. Computational methods emerge as a rapid and efficient solution to evaluate bird's-eye-view images taken by satellites, drones, or even street-view photos captured at the ground level of the urban scenery. Identifying unhealthy trees requires detecting the tree itself and its constituent parts to evaluate certain aspects that may indicate unhealthiness, being street-level images a cost-effective and feasible resource to support the fieldwork survey. This paper proposes detecting trees and their specific parts on street-view images through a Convolutional Neural Network model based on the well-known You Only Look Once network with a MobileNet as the backbone for feature extraction. Essentially, from a photo taken from the ground, the proposed method identifies trees, isolates them through their bounding boxes, identifies the crown and stem, and then estimates the height of the trees by using a specific handheld object as a reference in the images. Experiment results demonstrate the effectiveness of the proposed method.

This paper presents a Fuzzy Cognitive Map model to quantify implicit bias in structured datasets where features can be numeric or discrete. In our proposal, problem features are mapped to neural concepts that are initially activated by experts when running what-if simulations, whereas weights connecting the neural concepts represent absolute correlation/association patterns between features. In addition, we introduce a new reasoning mechanism equipped with a normalization-like transfer function that prevents neurons from saturating. Another advantage of this new reasoning mechanism is that it can easily be controlled by regulating nonlinearity when updating neurons’ activation values in each iteration. Finally, we study the convergence of our model and derive analytical conditions concerning the existence and unicity of fixed-point attractors.

Image restoration concerns mainly smoothing noise and de-blurring images that were corrupted either during acquisition or transmission. Since traditional deconvolution filters are highly dependent on specific kernels or prior knowledge to guide the deblurring process, image blur classification and further parameter estimation are critical for blind image de-blurring. This paper tackles the problem in three steps: (i) it first identifies the blur type for each input image, (ii) then it estimates the respective kernel parameter, and (iii) finally, it uses deconvolution filters to restore the blurred image. The proposed approach, called Deep Regressor Networks, showed promising results in general-purpose and remote sensing image datasets corrupted by different types and blur levels than some state-of-the-art techniques.

Research on remote sensing image classification significantly impacts essential human routine tasks such as urban planning and agriculture. Nowadays, the rapid advance in technology and the availability of many high-quality remote sensing images create a demand for reliable automation methods. The current paper proposes two novel deep learning-based architectures for image classification purposes, i.e., the Discriminant Deep Image Prior Network and the Discriminant Deep Image Prior Network+, which combine Deep Image Prior and Triplet Networks learning strategies. Experiments conducted over three well-known public remote sensing image datasets achieved state-of-the-art results, evidencing the effectiveness of using deep image priors for remote sensing image classification.

Breast cancer is among the most deadly diseases, distressing mostly women worldwide. Although traditional methods for detection have presented themselves as valid for the task, they still commonly present low accuracies and demand considerable time and effort from professionals. Therefore, a computer-aided diagnosis (CAD) system capable of providing early detection becomes hugely desirable. In the last decade, machine learning-based techniques have been of paramount importance in this context, since they are capable of extracting essential information from data and reasoning about it. However, such approaches still suffer from imbalanced data, specifically on medical issues, where the number of healthy people samples is, in general, considerably higher than the number of patients. Therefore this paper proposes the O2PF, a data oversampling method based on the unsupervised Optimum-Path Forest Algorithm. Experiments conducted over the full oversampling scenario state the robustness of the model, which is compared against three well-established oversampling methods considering three breast cancer and three general-purpose tasks for medical issues datasets.

Breast cancer is usually classified as either benign or malignant, where the former is not considered hazardous to health. Nonetheless, the benign tumors must be periodically monitored to control their activity and to prevent them from becoming malignant eventually. Several automated techniques have been proposed to aid the diagnosis by indicating potential tumor locations or by providing a broader insight. Although benign and malignant tumors are divided into four categories each, most of the works cope with their classification as just benign and malignant. This work addresses the problem of providing a more detailed classification of the tumors by proposing a deep-based architecture able to distinguish between eight types of tumors (i.e., four benign and four malignant). The proposed approach relies on the fusion of traditional convolution kernels with dilated convolutions before pooling, which can learn better spatial information, thus providing better feature detection prior to classification. Experimental results showed that the proposed approach outperformed the techniques compared in this work.

Restricted Boltzmann Machines (RBM) are stochastic neural networks mainly used for image reconstruction and unsupervised feature learning. An enhanced version, the temperature-based RBM (T-RBM), considers a new temperature parameter during the learning process that influences the neurons' activation. Nevertheless, the major vulnerability of such models concerns selecting an adequate system's temperature, which might lead them to inadequate training or even overfitting when wrongly set, thus limiting the network from predicting or working effectively over unseen data. This paper addresses the problem of selecting a suitable system's temperature through a meta-heuristic optimization process. Meta-heuristic-driven techniques, such as Particle Swarm Optimization, Bat Algorithm, and Artificial Bee Colony are employed to find proper values for the temperature parameter. Additionally, for comparison purposes, three standard temperature values and a random search are used as baselines. The results revealed that optimizing T-RBM is suitable for training purposes, primarily due to their complex fitness landscape, which makes fine-tuning temperatures a nontrivial task.

In the last century, Albert Einstein's perceptions of the world afforded a revolution in the understanding of the universe. In his theory of general relativity, he describes the space-time continuum, a concept capable of explaining several phenomena, ranging from gravity to black holes and supernovas. Further, it also provides a set of formulations to generalize classical physics concepts to accommodate the relativistic notions. Meanwhile, several mathematicians have been working on optimization tools aiming to solve complex problems associated with a large number of variables. Nowadays, despite the computational power, many daily tasks still pose a challenge and are becoming more prohibitives, mostly due to the massive amount of data to be processed. Therefore, efficient optimization techniques are more desirable than ever. In this context, metaheuristic optimization has arisen, i.e., stochastic nature-inspired methods capable of finding sub-optimal solutions for complex problems with a reasonable computational effort. However, such approaches still suffer from some drawbacks related to low convergence and getting stuck on local optima, among others. Therefore, in this paper, we introduce relativistic concepts into the well-known meta-heuristic optimization technique Particle Swarm optimization (PSO). The experimental results evince the robustness of the proposed approach compared to the standard PSO as well as three other variations for five benchmarking functions.

Convolutional Neural Networks (CNN) have been being widely employed to solve the challenging remote sensing task of aerial scene classification. Nevertheless, it is not straightforward to find single CNN models that can solve all aerial scene classification tasks, allowing the development of a better alternative, which is to fuse CNN-based classifiers into an ensemble. However, an appropriate choice of the classifiers that will belong to the ensemble is a critical factor, as it is unfeasible to employ all the possible classifiers in the literature. Therefore, this work proposes a novel framework based on meta-heuristic optimization for creating optimized ensembles in the context of aerial scene classification. The experimental results were performed across nine meta-heuristic algorithms and three aerial scene literature datasets, being compared in terms of effectiveness (accuracy), efficiency (execution time), and behavioral performance in different scenarios. Our results suggest that the Univariate Marginal Distribution Algorithm shows more effective and efficient results than other commonly used meta-heuristic algorithms, such as Genetic Programming and Particle Swarm Optimization.

Machine learning approaches have achieved remarkable advances over the last decades, especially in supervised learning tasks such as classification. Meanwhile, multimedia data and applications experienced an explosive growth, becoming ubiquitous in diverse domains. Due to the huge increase in multimedia data collections and the lack of labeled data in several scenarios, creating methods capable of exploiting the unlabeled data and operating under weakly supervision is imperative. In this work, we propose a rank-based model to exploit contextual information encoded in the unlabeled data in order to perform weakly supervised classification. We employ different rank-based correlation measures for identifying strong similarities relationships and expanding the labeled set in an unsupervised way. Subsequently, the extended labeled set is used by a classifier to achieve better accuracy results. The proposed weakly supervised approach was evaluated on multimedia classification tasks, considering several combinations of rank correlation measures and classifiers. An experimental evaluation was conducted on 4 public image datasets and different features. Very positive gains were achieved in comparison with various semi-supervised and supervised classifiers taken as baselines when considering the same amount of labeled data.

Different techniques have emerged in the deep learning scenario, such as Convolutional Neural Networks, Deep Belief Networks, and Long Short-Term Memory Networks, to cite a few. In lockstep, regularization methods, which aim to prevent overfitting by penalizing the weight connections, or turning off some units, have been widely studied either. In this paper, we present a novel approach called MaxDropout, a regularizer for deep neural network models that works in a supervised fashion by removing (shutting off) the prominent neurons (i.e., most active) in each hidden layer. The model forces fewer activated units to learn more representative information, thus providing sparsity. Regarding the experiments, we show that it is possible to improve existing neural networks and provide better results in neural networks when Dropout is replaced by MaxDropout. The proposed method was evaluated in image classification, achieving comparable results to existing regularizers, such as Cutout and RandomErasing, also improving the accuracy of neural networks that uses Dropout by replacing the existing layer by MaxDropout.

The task of learning to rank has been widely studied by the machine learning community, mainly due to its use and great importance in information retrieval, data mining, and natural language processing. Therefore, ranking accurately and learning to rank are crucial tasks. Context-Based Information Retrieval systems have been of great importance to reduce the effort of finding relevant data. Such systems have evolved by using machine learning techniques to improve their results, but they are mainly dependent on user feedback. Although information retrieval has been addressed in different works along with classifiers based on Optimum-Path Forest (OPF), these have so far not been applied to the learning to rank task. Therefore, the main contribution of this work is to evaluate classifiers based on Optimum-Path Forest, in such a context. Experiments were performed considering the image retrieval and ranking scenarios, and the performance of OPF-based approaches was compared to the well-known SVM-Rank pairwise technique and a baseline based on distance calculation. The experiments showed competitive results concerning precision and outperformed traditional techniques in terms of computational load.

Image segmentation is the task of assigning a label to each image pixel. When the number of labels is greater than two (multi-label) the segmentation can be modelled as a multi-cut problem in graphs. In the general case, finding the minimum cut in a graph is an NP-hard problem, in which improving the results concerning time and quality is a major challenge. This paper addresses the multi-label problem applied in interactive image segmentation. The proposed approach makes use of dynamic programming to initialize an α-expansion, thus reducing its runtime, while keeping the Dice-score measure in an interactive segmentation task. Over BSDS data set, the proposed algorithm was approximately 51.2% faster than its standard counterpart, 36.2% faster than Fast Primal-Dual (FastPD) and 10.5 times faster than quadratic pseudo-boolean optimization (QBPO) optimizers, while preserving the same segmentation quality.

Abstract In the Big Data Era, recommender systems perform a fundamental role in data management and information filtering. In this context, Collaborative Filtering (CF) persists as one of the most prominent strategies to effectively deal with large datasets and is capable of offering users interesting content in a recommendation fashion. Nevertheless, it is well‐known CF recommenders suffer from data sparsity, mainly in cold‐start scenarios, substantially reducing the quality of recommendations. In the vast literature about the aforementioned topic, there are numerous solutions, in which the state‐of‐the‐art contributions are, in some sense, conditioned or associated with traditional CF methods such as Matrix Factorization (MF), that is, they rely on linear optimization procedures to model users and items into low‐dimensional embeddings. To overcome the aforementioned challenges, there has been an increasing number of studies exploring deep learning techniques in the CF context for latent factor modelling. In this research, authors conduct a systematic review focusing on state‐of‐the‐art literature on deep learning techniques applied in collaborative filtering recommendation, and also featuring primary studies related to mitigating the cold start problem. Additionally, authors considered the diverse non‐linear modelling strategies to deal with rating data and side information, the combination of deep learning techniques with traditional CF‐based linear methods, and an overview of the most used public datasets and evaluation metrics concerning CF scenarios.

In the past decades, fuzzy logic has played an essential role in many research areas. Alongside, graph-based pattern recognition has shown to be of great importance due to its flexibility in partitioning the feature space using the background from graph theory. Some years ago, a new framework for supervised, semisupervised, and unsupervised learning, named optimum-path forest (OPF), was proposed with competitive results in several applications, besides comprising a low computational burden. In this article, we propose the fuzzy OPF, an improved version of the standard OPF classifier, that learns the samples' membership in an unsupervised fashion, which are further incorporated during supervised training. Such information is used to identify the most relevant training samples, thus improving the classification step. Experiments conducted over 12 public datasets highlight the robustness of the proposed approach, which behaves similarly to standard OPF in worst case scenarios.

Fitness landscape has been one of the main limitations regarding optimization tasks. Although meta-heuristic techniques have achieved outstanding results over a large variety of problems, some issues related to the function geometry and the risk to get trapped from local optima are issues that still require attention. To deal with this problem, we propose the Quaternion-based Backtracking Search Optimization Algorithm, a variant of the standard Backtracking Search Optimization Algorithm that maps each decision variable in a tensor onto a hypercomplex search space, whose landscape is expected to be smoother. Experiments conducted using nine benchmarking functions showed considerably better results than the ones achieved over standard search spaces, as well as more accurate results than some quaternion-based methods as well.

Multiple-instance (MI) learning aims at modeling problems that are better described by several instances of a given sample instead of individual descriptions often employed by standard machine learning approaches. In binary-driven MI problems, the entire bag is considered positive if one (at least) sample is labeled as positive. On the other hand, a bag is considered negative if it contains all samples labeled as negative as well. In this paper, we introduced the Optimum-Path Forest (OPF) classifier to the context of multiple-instance learning paradigm, and we evaluated it in different scenarios that range from molecule description, text categorization, and anomaly detection in well-drilling report classification. The experimental results showed that two different OPF classifiers are very much suitable to handle problems in the multiple-instance learning paradigm.

Restricted Boltzmann Machines achieved notorious popularity in the scientific community in the last decade due to outstanding results in a wide range of applications and also for providing the required mechanisms to build successful deep learning models, i.e., Deep Belief Networks and Deep Boltzmann Machines. However, their main bottleneck is related to the learning step, which is usually time-consuming. In this paper, we introduce a Sigmoid-like family of functions based on the Kaniadakis entropy formulation in the context of the RBM learning procedure. Experiments concerning binary image reconstruction are conducted in four public datasets to evaluate the robustness of the proposed approach. The results suggest that such a family of functions is suitable to increase the convergence rate when compared to standard functions employed by the research community.

Since the past decades, the world has been transformed into a digital society, where every individual is living with a unique identifier. The primary purpose of this id is to distinguish from others and to deal with digital machines which are surrounding the world. Recently, many researchers showed that the brain electrical activity or electroencephalogram (EEG) signals could provide robust and unique features that can be considered as a new biometric authentication technique, given that accurately methods to decompose the signals must also be considered. This paper proposes a novel method for EEG signal denoising based on the multi-objective Flower Pollination Algorithm and the Wavelet Transform (MOFPA-WT) to extract useful features from denoised signals. MOFPA-WT is tested using a standard EEG signal dataset, namely, EEG motor movement/imagery dataset, and its performance is evaluated using three criteria: (i) accuracy, (ii) true acceptance rate, and (iii) false acceptance rate. We show that the proposed method can achieve results that are comparable to the state-of-the-art ones, as well as we draw future directions towards the research area.

Appropriate metrics are paramount for machine learning and pattern recognition. In Content-based Image Retrieval-oriented applications, low-level features and pairwise-distance metrics are usually not capable of representing similarity among the objects as observed by humans. Therefore, metric learning from available data has become crucial in such applications, but just a few related approaches take into account the contextual information inherent from the samples for a better accuracy performance. In this paper, we propose a novel approach which combines an unsupervised manifold learning algorithm with the Optimum-Path Forest (OPF) classifier to obtain more accurate recognition rates, as well as we show it can outperform standard OPF-based classifiers that are trained over the original manifold. Experiments conducted in some public datasets evidenced the validity of metric learning in the context of OPF classifiers.

Satellite images have been used in a number of applications, both in the academy and in the industry. One critical purpose concerns the land-use classification, which aims at automatically identifying different land-use applications, which range from economy and environmental monitoring to resources planning. In this paper, we introduce a new machine learning technique called Finite Element Machines (FEMa) in the context of land-use classification using satellite images. We show that FEMa can obtain results that are comparable to some state-of-the-art techniques in the literature.

Recently, drone images have been used in a number of applications, mainly for pollution control and surveillance purposes. In this paper, we introduce the well-known Convolutional Neural Networks in the context of environmental monitoring using drone images, and we show their robustness in real-world images obtained from uncontrolled scenarios. We consider a transfer learning-based approach and compare two neural models, i.e., VGG16 and VGG19, to distinguish four classes: “water”, “deforesting area”, “forest”, and “buildings”. The results are analyzed by experts in the field and considered pretty much reasonable.

Stroke is an injury that affects the brain tissue, mainly caused by changes in the blood supply to a particular region of the brain. As consequence, some specific functions related to that affected region can be reduced, decreasing the quality of life of the patient. In this work, we deal with the problem of stroke detection in Computed Tomography (CT) images using Convolutional Neural Networks (CNN) optimized by Particle Swarm optimization (PSO). We considered two different kinds of strokes, ischemic and hemorrhagic, as well as making available a public dataset to foster the research related to stroke detection in the human brain. The dataset comprises three different types of images for each case, i.e., the original CT image, one with the segmented cranium and an additional one with the radiological density's map. The results evidenced that CNN's are suitable to deal with stroke detection, obtaining promising results.

Well drilling monitoring is an essential task to prevent faults, save resources, and take care of environmental and eco-planning businesses. During drilling, it is required that staff fill out a log to keep track of the activities that are currently occurring. With such data analyzed and processed, it is possible to learn how to prevent faults and take corrective actions in realtime. However, the most important information is usually stored in a free-text format, thus complicating the task of automated text mining. In this work, we introduce the Optimum-Path Forest (OPF) for sentence classification in drilling reports and compare its results against some state-of-art results. We show that OPF combined with text-based features are a compelling source to learn patterns in drilling reports.

Machine learning techniques have been actively pursued in the last years, mainly due to the great number of applications that make use of some sort of intelligent mechanism for decision-making processes. In this context, we shall highlight pruning strategies, which provide heuristics to select from a collection of classifiers the ones that can really improve recognition rates when working together. In this paper, we present an ensemble pruning approach of Optimum-Path Forest classifiers based on metaheuristics, as well as we introduced the concept of quaternions in ensemble pruning strategies. Experimental results over synthetic and real datasets showed the effectiveness and efficiency of the proposed approach for classification problems.

Regression-based tasks have been the forerunner regarding the application of machine learning tools in the context of data mining. Problems related to price and stock prediction, selling estimation, and weather forecasting are commonly used as benchmarking for the comparison of regression techniques, just to name a few. Neural Networks, Decision Trees and Support Vector Machines are the most widely used approaches concerning regression-oriented applications, since they can generalize well in a number of different applications. In this work, we propose an efficient and effective regression technique based on the Finite Element Method (FEM) theory, hereinafter called Finite Element Machine for Regression (FEMaR). The proposed approach has only one parameter and it has a quadratic complexity for both training and classification phases when we use basis functions that obey some properties, as well as we show the proposed approach can obtain very competitive results when compared against some state-of-the-art regression techniques.

Biometric systems present some important advantages over the traditional knowledge-or possess-oriented identification systems, such as a guarantee of authenticity and convenience. However, due to their widespread usage in our society and despite the difficulty in attacking them, nowadays criminals are already developing techniques to simulate physical, physiological and behavioral traits of valid users, the so-called spoofing attacks. In this sense, new countermeasures must be developed and integrated with the traditional biometric systems to prevent such frauds. In this work, we present a novel robust and efficient approach to detect spoofing attacks in biometric systems (fingerprint-based ones) using a deep learning-based model: the Deep Boltzmann Machine (DBM). By extracting and working with high-level features from the original data, DBM can deal with complex patterns and work with features that can not be easily forged. The results show the proposed approach outperforms other state-of-the-art techniques, presenting high accuracy in terms of attack detection and allowing working with less labeled data.

Abstract Recently, multi‐ and many‐objective meta‐heuristic algorithms have received considerable attention due to their capability to solve optimization problems that require more than one fitness function. This paper presents a comprehensive study of these techniques applied in the context of machine learning problems. Three different topics are reviewed in this work: (a) feature extraction and selection, (b) hyper‐parameter optimization and model selection in the context of supervised learning, and (c) clustering or unsupervised learning. The survey also highlights future research towards related areas.

Contextual image classification aims at considering the information about nearby samples in the learning process in order to provide more accurate results. In this paper, we propose a locally-adaptive Optimum-Path Forest classifier together with Markov Random Fields (MRF) that surpasses its naïve version, which was recently presented in the literature. The experimental results over four satellite images demonstrated the proposed approach an outperform previous results, as well as it can perform MRF parameter learning much faster than its former version.

Contextual-based learning aims at considering neighboring pixels to improve pixelwise-oriented classification techniques. In this letter, we presented a metaheuristic framework for the optimization of nondiscrete Markovian models considering the optimum-path forest (OPF) classifier, and we proposed a postprocessing procedure to avoid overcorrection over high-frequency regions. The proposed approach outperformed previous results obtained with standard OPF in satellite imagery.

Parkinson's disease (PD) has affected millions of people world-wide, being its major problem the loss of movements and, consequently, the ability of working and locomotion. Although we can find several works that attempt at dealing with this problem out there, most of them make use of datasets composed by a few subjects only. In this work, we present some results toward the automated diagnosis of PD by means of computer vision-based techniques in a dataset composed by dozens of patients, which is one of the main contributions of this work. The dataset is part of a joint research project that aims at extracting both visual and signal-based information from healthy and PD patients in order to go forward the early diagnosis of PD patients. The dataset is composed by handwriting clinical exams that are analyzed by means of image processing and machine learning techniques, being the preliminary results encouraging and promising. Additionally, a new quantitative feature to measure the amount of tremor of an individual's handwritten trace called Mean Relative Tremor is also presented.

Computer-Aided Diagnosis (CAD) can be divided into two main categories: CADe (Computer-Aided Detection), which is focused on the detection of structures of interest, as well as to assist radiologists to find out signals of interest that might be hidden to human vision, and the CADx (Computer-Aided Diagnosis), which works as a second observer, being responsible to give an opinion on a specific lesion. In CADe - based systems, the identification of mammograms with and without masses is highly needed to reduce the false positive rates regarding the automatic selection of regions of interest. The main contribution of this study is to introduce the unsupervised classifier Optimum-Path Forest to identify breast masses, and to evaluate its performance against with two other unsupervised techniques (Gaussian Mixture Model and k-Means) using texture features from images obtained from a private dataset composed by 120 images with and without the presence of masses.

Nowadays, people use smartphones and tablets with the very same purposes as desktop computers: web browsing, social networking and home-banking, just to name a few. However, we are often facing the problem of keeping our information protected and trustworthy. As a result of their popularity and functionality, mobile devices are a growing target for malicious activities. In such context, mobile malwares have gained significant ground since the emergence and growth of smartphones and handheld devices, becoming a real threat. In this paper, we introduced a recently developed pattern recognition technique called Optimum-Path Forest in the context of malware detection, as well we present "DroidWare", a new public dataset to foster the research on mobile malware detection. In addition, we also proposed to use Restricted Boltzmann Machines for unsupervised feature learning in the context of malware identification.

In the past years, SMS messages have shown to be a profitable revenue to the cell-phone industries, being one of the most used communication systems to date. However, this very same scenario has led spammers to concentrate their attentions into spreading spam messages through SMS, thus achieving some success due to the lack of proper tools to cope with this problem. In this paper, we introduced the Optimum-Path Forest classifier to the context of spam filtering in SMS messages, as well as we compared it against with some state-of-the-art supervised pattern recognition techniques. We have shown promising results with an user-friendly classifier, which requires minimum user interaction and less knowledge about the dataset.

Unsupervised land-cover classification aims at learning intrinsic properties of spectral and spatial features for the task of area coverage in urban and rural areas. In this paper, we propose to model the problem of optimizing the well-known k-means algorithm by combining different variations of the Harmony Search technique using Genetic Programming (GP). We have shown GP can improve the recognition rates when using one optimization technique only, but it still deserves a deeper study when we have a very good individual technique to be combined.

Sequential learning-based pattern classification aims at providing more accurate labeled maps by adding an extra step of classification using an augmented feature vector. In this paper, we evaluated the robustness of Optimum-Path Forest (OPF) classifier in the context of land-cover classification using both satellite and radar images, showing OPF can benefit from sequential learning theoretical basis.

Evolutionary algorithms have been widely used for Artificial Neural Networks (ANN) training, being the idea to update the neurons' weights using social dynamics of living organisms in order to decrease the classification error. In this paper, we have introduced Social-Spider Optimization to improve the training phase of ANN with Multilayer perceptrons, and we validated the proposed approach in the context of Parkinson's Disease recognition. The experimental section has been carried out against with five other well-known meta-heuristics techniques, and it has shown SSO can be a suitable approach for ANN-MLP training step.

In Computer-Aided Diagnosis-based schemes in mammography analysis each module is interconnected, which directly affects the system operation as a whole. The identification of mammograms with and without masses is highly needed to reduce the false positive rates regarding the automatic selection of regions of interest for further image segmentation. This study aims to evaluate the performance of three techniques in classifying regions of interest as containing masses or without masses (without clinical findings), as well as the main contribution of this work is to introduce the Optimum-Path Forest (OPF) classifier in this context, which has never been done so far. Thus, we have compared OPF against with two sorts of neural networks in a private dataset composed by 120 images: Radial Basis Function and Multilayer Perceptron (MLP). Texture features have been used for such purpose, and the experiments have demonstrated that MLP networks have been slightly better than OPF, but the former is much faster, which can be a suitable tool for real-time recognition systems.

Meta-heuristic-based feature selection has been paramount in the last years, mainly because of its simplicity, effectiveness and also efficiency in some cases. Such approaches are based on the social dynamics of living organisms, and can vary from birds, bees, bats and ants. Very recently, an optimization algorithm based on krill herd (KH) was proposed for continuous-valued applications, and it has been more accurate than some state-of-the-art techniques. In this paper, we propose a binary optimization version of KH technique, and we validate it for feature selection purposes in several datasets. The experiments showed the proposed technique outperforms three other meta-heuristic-based approaches for this task, being also so fast as the compared techniques.

In this paper, we show how to improve the Radial Basis Function Neural Networks effectiveness by using the Optimum-Path Forest clustering algorithm, since it computes the number of clusters on-the-fly, which can be very interesting for finding the Gaussians that cover the feature space. Some commonly used approaches for this task, such as the well-known fc-means, require the number of classes/clusters previous its performance. Although the number of classes is known in supervised applications, the real number of clusters is extremely hard to figure out, since one class may be represented by more than one cluster. Experiments over 9 datasets together with statistical analysis have shown the suitability of OPF clustering for the RBF training step.

The choice of hyper-parameters in Support Vector Machines (SVM)-based learning is a crucial task, since different values may degrade its performance, as well as can increase the computational burden. In this paper, we introduce a recently developed nature-inspired optimization algorithm to find out suitable values for SVM kernel mapping named Social-Spider Optimization (SSO). We compare the results obtained by SSO against with a Grid-Search, Particle Swarm Optimization and Harmonic Search. Statistical evaluation has showed SSO can outperform the compared techniques for some sort of kernels and datasets.

Although hyperspectral images acquired by on-board satellites provide information from a wide range of wavelengths in the spectrum, the obtained information is usually highly correlated. This paper proposes a novel framework to reduce the computation cost for large amounts of data based on the efficiency of the optimum-path forest (OPF) classifier and the power of metaheuristic algorithms to solve combinatorial optimizations. Simulations on two public data sets have shown that the proposed framework can indeed improve the effectiveness of the OPF and considerably reduce data storage costs.

Land cover classification has been paramount in the last years. Since the amount of information acquired by satellite on-board imaging systems has increased, there is a need for automatic tools that can tackle such problem. Despite the fact that one can find several works in the literature, we propose a novel methodology for land cover classification by means of the optimum-path forest (OPF) framework, which has never been applied to this context up to date. Experiments were conducted in supervised and unsupervised situations against some state-of-the-art pattern recognition techniques, such as support vector machines, Bayesian classifier, k-means, and mean shift. We had shown that supervised OPF can outperform such approaches, being much faster than all. In regard to clustering techniques, all classifiers have achieved similar results.

Image restoration is a research field that attempts to recover a blurred and noisy image. Since it can be modeled as a linear system, we propose in this paper to use the meta-heuristics optimization algorithm Harmony Search (HS) to find out near-optimal solutions in a Projections Onto Convex Sets-based formulation to solve this problem. The experiments using HS and four of its variants have shown that we can obtain near-optimal and faster restored images than other evolutionary optimization approach.

Feature selection has been actively pursued in the last years, since to find the most discriminative set of features can enhance the recognition rates and also to make feature extraction faster. In this paper, the propose a new feature selection called Binary Cuckoo Search, which is based on the behavior of cuckoo birds. The experiments were carried out in the context of theft detection in power distribution systems in two datasets obtained from a Brazilian electrical power company, and have demonstrated the robustness of the proposed technique against with several others nature-inspired optimization techniques.

Image categorization by means of bag of visual words has received increasing attention by the image processing and vision communities in the last years. In these approaches, each image is represented by invariant points of interest which are mapped to a Hilbert Space representing a visual dictionary which aims at comprising the most discriminative features in a set of images. Notwithstanding, the main problem of such approaches is to find a compact and representative dictionary. Finding such representative dictionary automatically with no user intervention is an even more difficult task. In this paper, we propose a method to automatically find such dictionary by employing a recent developed graph-based clustering algorithm called Optimum-Path Forest, which does not make any assumption about the visual dictionary's size and is more efficient and effective than the state-of-the-art techniques used for dictionary generation.

In this paper we addressed the problem of dimensionality reduction in hyperspectral imagery classification by combining OPF classifier together with three recent evolutionary-based optimization algorithms: PSO, HS and GSA. We conducted experiments with two public datasets (Indian Pines and Salinas), which demonstrated that OPF combined with HS and GSA have obtained promising results, being the former the fastest approach. In regard to Indian Pines dataset, HS and GSA have achieved close classification rates, but HS has selected 46.25% less bands, which means a faster feature extraction step. For future works, we intend to provide a more detailed convergence analysis for PSO, HS and GSA, and also to introduce novel evolutionary-based band selection techniques and also to apply these methodologies for hyperspectral image classification in forest and agriculture applications.

In this paper we shed light over the problem of landslide automatic recognition using supervised classification, and we also introduced the OPF classifier in this context. We employed two images acquired from Geoeye-MS satellite at March-2010 in the northwest (high steep areas) and north sides (pipeline area) covering the area of Duque de Caxias city, Rio de Janeiro State, Brazil. The landslide recognition rate has been assessed through a cross-validation with 10 runnings. In regard to the classifiers, we have used OPF against SVM with Radial Basis Function for kernel mapping and a Bayesian classifier. We can conclude that OPF, Bayes and SVM achieved high recognition rates, being OPF the fastest approach.

Majority of biometric researchers focus on the accuracy of matching using biometrics databases, including iris databases, while the scalability and speed issues have been neglected. In the applications such as identification in airports and borders, it is critical for the identification system to have low-time response. In this paper, a graph-based framework for pattern recognition, called Optimum-Path Forest (OPF), is utilized as a classifier in a pre-developed iris recognition system. The aim of this paper is to verify the effectiveness of OPF in the field of iris recognition, and its performance for various scale iris databases. This paper investigates several classifiers, which are widely used in iris recognition papers, and the response time along with accuracy. The existing Gauss-Laguerre Wavelet based iris coding scheme, which shows perfect discrimination with rotary Hamming distance classifier, is used for iris coding. The performance of classifiers is compared using small, medium, and large scale databases. Such comparison shows that OPF has faster response for large scale database, thus performing better than more accurate but slower Bayesian classifier.

Nowadays, organizations face the problem of keeping their information protected, available and trustworthy. In this context, machine learning techniques have also been extensively applied to this task. Since manual labeling is very expensive, several works attempt to handle intrusion detection with traditional clustering algorithms. In this paper, we introduce a new pattern recognition technique called Optimum-Path Forest (OPF) clustering to this task. Experiments on three public datasets have showed that OPF classifier may be a suitable tool to detect intrusions on computer networks, since it outperformed some state-of-the-art unsupervised techniques.

With several good research groups actively working in machine learning (ML) approaches, we have now the concept of self-containing machine learning solutions that oftentimes work out-of-the-box leading to the concept of ML black-boxes. Although it is important to have such black-boxes helping researchers to deal with several problems nowadays, it comes with an inherent problem increasingly more evident: we have observed that researchers and students are progressively relying on ML black-boxes and, usually, achieving results without knowing the machinery of the classifiers. In this regard, this paper discusses the use of machine learning black-boxes and poses the question of how far we can get using these out-of-the-box solutions instead of going deeper into the machinery of the classifiers. The paper focuses on three aspects of classifiers: (1) the way they compare examples in the feature space; (2) the impact of using features with variable dimensionality; and (3) the impact of using binary classifiers to solve a multi-class problem. We show how knowledge about the classifier's machinery can improve the results way beyond out-of-the-box machine learning solutions.

In this paper we deal with the problem of feature selection by introducing a new approach based on Gravitational Search Algorithm (GSA). The proposed algorithm combines the optimization behavior of GSA together with the speed of Optimum-Path Forest (OPF) classifier in order to provide a fast and accurate framework for feature selection. Experiments on datasets obtained from a wide range of applications, such as vowel recognition, image classification and fraud detection in power distribution systems are conducted in order to asses the robustness of the proposed technique against Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA) and a Particle Swarm Optimization (PSO)-based algorithm for feature selection.

Different from the first attempts to solve the image categorization problem (often based on global features), recently, several researchers have been tackling this research branch through a new vantage point - using features around locally invariant interest points and visual dictionaries. Although several advances have been done in the visual dictionaries literature in the past few years, a problem we still need to cope with is calculation of the number of representative words in the dictionary. Therefore, in this paper we introduce a new solution for automatically finding the number of visual words in an N-Way image categorization problem by means of supervised pattern classification based on optimum-path forest.

Land use classification has been paramount in the last years, since we can identify illegal land use and also to monitor deforesting areas. Although one can find several research works in the literature that address this problem, we propose here the land use recognition by means of Optimum-Path Forest Clustering (OPF), which has never been applied to this context up to date. Experiments among Optimum-Path Forest, Mean Shift and K-Means demonstrated the robustness of OPF for automatic land use classification of images obtained by CBERS-2B and Ikonos-2 satellites.

In this paper we would like to shed light the problem of efficiency and effectiveness of image classification in large datasets. As the amount of data to be processed and further classified has increased in the last years, there is a need for faster and more precise pattern recognition algorithms in or- der to perform online and offline training and classification procedures. We deal here with the problem of moist area classification in radar image in a fast manner. Experimental results using Optimum-Path Forest and its training set pruning algorithm also provided and discussed.

Although non-technical losses automatic identification has been massively studied, the problem of selecting the most representative features in order to boost the identification accuracy has not attracted much attention in this context. In this paper, we focus on this problem applying a novel feature selection algorithm based on Particle Swarm Optimization and Optimum-Path Forest. The results demonstrated that this method can improve the classification accuracy of possible frauds up to 49% in some datasets composed by industrial and commercial profiles.

Intrusion detection systems that make use of artificial intelligence techniques in order to improve effectiveness have been actively pursued in the last decade. Neural networks and Support Vector Machines have been also extensively applied to this task. However, their complexity to learn new attacks has become very expensive, making them inviable for a real time retraining. In this research, we introduce a new pattern classifier named Optimum-Path Forest (OPF) to this task, which has demonstrated to be similar to the state-of-the-art pattern recognition techniques, but extremely more efficient for training patterns. Experiments on public datasets showed that OPF classifier may be a suitable tool to detect intrusions on computer networks, as well as allow the algorithm to learn new attacks faster than the other techniques.

The applications of Automatic Vowel Recognition (AVR), which is a sub-part of fundamental importance in most of the speech processing systems, vary from automatic interpretation of spoken language to biometrics. State-of-the-art systems for AVR are based on traditional machine learning models such as Artificial Neural Networks (ANNs) and Support Vector Machines (SVMs), however, such classifiers can not deal with efficiency and effectiveness at the same time, existing a gap to be explored when real-time processing is required. In this work, we present an algorithm for AVR based on the Optimum-Path Forest (OPF), which is an emergent pattern recognition technique recently introduced in literature. Adopting a supervised training procedure and using speech tags from two public datasets, we observed that OPF has outperformed ANNs, SVMs, plus other classifiers, in terms of training time and accuracy.

Traditional pattern recognition techniques can not handle the classification of large datasets with both efficiency and effectiveness. In this context, the Optimum-Path Forest (OPF) classifier was recently introduced, trying to achieve high recognition rates and low computational cost. Although OPF was much faster than Support Vector Machines for training, it was slightly slower for classification. In this paper, we present the Efficient OPF (EOPF), which is an enhanced and faster version of the traditional OPF, and validate it for the automatic recognition of white matter and gray matter in magnetic resonance images of the human brain.

The decision correctness in expert systems strongly depends on the accuracy of a pattern classifier, whose learning is performed from labeled training samples. Some systems, however, have to manage, store, and process a large amount of data, making also the computational efficiency of the classifier an important requirement. Examples are expert systems based on image analysis for medical diagnosis and weather forecasting. The learning time of any pattern classifier increases with the training set size, and this might be necessary to improve accuracy. However, the problem is more critical for some popular methods, such as artificial neural networks and support vector machines (SVM), than for a recently proposed approach, the optimum-path forest (OPF) classifier. In this letter, we go beyond by presenting a robust approach to reduce the training set size and still preserve good accuracy in OPF classification. We validate the method using some data sets and for rainfall occurrence estimation based on satellite image analysis. The experiments use SVM and OPF without pruning of training patterns as baselines.