Open Access
Additional feature layers from ordered aggregations for deep neural networks
(IEEE, 2020) Domínguez Catena, Iris; Paternain Dallo, Daniel; Galar Idoate, Mikel; Institute of Smart Cities - ISC; Universidad Pública de Navarra / Nafarroako Unibertsitate Publikoa
In the last years we have seen huge advancements in the area of Machine Learning, specially with the use of Deep Neural Networks. One of the most relevant examples is in image classification, where convolutional neural networks have shown to be a vital tool, hard to replace with any other techniques. Although aggregation functions, such as OWA operators, have been previously used on top of neural networks, usually to aggregate the outputs of different networks or systems (ensembles), in this paper we propose and explore a new way of using OWA aggregations in deep learning. We implement OWA aggregations as a new layer inside a convolutional neural network. These layers are used to learn additional order-based information from the feature maps of a certain layer, and then the newly generated information is used as a complement input for the following layers. We carry out several tests introducing the new layer in a VGG13-based reference network and show that this layer introduces new knowledge into the network without substantially increasing training times.
Open Access
Enhancing DreamBooth with LoRA for generating unlimited characters with stable diffusion
(IEEE, 2024-09-09) Pascual Casas, Rubén; Maiza Coupin, Adrián Mikel; Sesma Sara, Mikel; Paternain Dallo, Daniel; Galar Idoate, Mikel; Estadística, Informática y Matemáticas; Estatistika, Informatika eta Matematika; Institute of Smart Cities - ISC; Universidad Pública de Navarra / Nafarroako Unibertsitate Publikoa, PJUPNA2023-11377
This paper addresses the challenge of generating unlimited new and distinct characters that encompass the style and shared visual characteristics of a limited set of human designed characters. This is a relevant problem in the audiovisual industry, as the ability to rapidly produce original characters that adhere to specific characteristics greatly increases the possibilities in the production of movies, series, or video games. Our solution is built upon DreamBooth, a widely extended fine-tuning method for text-to-image models. We propose an adaptation focusing on two main challenges: the impracticality of relying on detailed image prompts for character description and the few-shot learning scenario with a limited set of characters available for training. To solve these issues, we introduce additional character-specific tokens to DreamBooth training and remove its class-specific regularization dataset. For an unlimited generation of characters, we propose the usage of random tokens and random embeddings. This proposal is tested on two specialized datasets and the results shows our method¿s capability to produce diverse characters that adhere to a style and visual characteristics. An ablation study to analyze the contributions of the proposed modifications is also developed.
Open Access
Diffusion models for remote sensing imagery semantic segmentation
(IEEE, 2023-10-20) Ayala Lauroba, Christian; Sesma Redín, Rubén; Aranda, Carlos; Galar Idoate, Mikel; Institute of Smart Cities - ISC; Universidad Pública de Navarra / Nafarroako Unibertsitate Publikoa, PJUPNA25-2022
Denoising Diffusion Probabilistic Models have exhibited impressive performance for generative modelling of images. This paper aims to explore the potential of diffusion models for semantic segmentation tasks in the context of remote sensing. The major challenge of employing these models for semantic segmentation tasks is the generative nature of the model, which produces an arbitrary segmentation mask from a random noise input. Therefore, the diffusion process needs to be constrained to produce a segmentation mask that matches the target image. To address this issue, the denoising process is conditioned by utilizing the input image as a reference. In the experimental study, the proposed model is compared against other state-of-the-art semantic segmentation architectures using the Massachusetts Buildings Aerial dataset. The results of this study provide valuable insights into the potential of diffusion models for semantic segmentation tasks in the field of remote sensing.
Open Access
Dissimilarity based choquet integrals
(Springer, 2020) Bustince Sola, Humberto; Mesiar, Radko; Fernández Fernández, Francisco Javier; Galar Idoate, Mikel; Paternain Dallo, Daniel; Estadística, Informática y Matemáticas; Estatistika, Informatika eta Matematika
In this paper, in order to generalize the Choquet integral, we replace the difference between inputs in its definition by a restricted dissimilarity function and refer to the obtained function as d-Choquet integral. For some particular restricted dissimilarity function the corresponding d-Choquet integral with respect to a fuzzy measure is just the ‘standard’ Choquet integral with respect to the same fuzzy measure. Hence, the class of all d-Choquet integrals encompasses the class of all 'standard' Choquet integrals. This approach allows us to construct a wide class of new functions, d-Choquet integrals, that are possibly, unlike the 'standard' Choquet integral, outside of the scope of aggregation functions since the monotonicity is, for some restricted dissimilarity function, violated and also the range of such functions can be wider than [0, 1], in particular it can be [0, n].
Open Access
Addressing the overlapping data problem in classification using the one-vs-one decomposition strategy
(IEEE, 2019) Sáez, José Antonio; Galar Idoate, Mikel; Krawczyk, Bartosz; Institute of Smart Cities - ISC
Learning good-performing classifiers from data with easily separable classes is not usually a difficult task for most of the algorithms. However, problems affecting classifier performance may arise when samples from different classes share similar characteristics or are overlapped, since the boundaries of each class may not be clearly defined. In order to address this problem, the majority of existing works in the literature propose to either adapt well-known algorithms to reduce the negative impact of overlapping or modify the original data by introducing/removing features which decrease the overlapping region. However, these approaches may present some drawbacks: the changes in specific algorithms may not be useful for other methods and modifying the original data can produce variable results depending on data characteristics and the technique used later. An unexplored and interesting research line to deal with the overlapping phenomenon consists of decomposing the problem into several binary subproblems to reduce its complexity, diminishing the negative effects of overlapping. Based on this novel idea in the field of overlapping data, this paper proposes the usage of the One-vs-One (OVO) strategy to alleviate the presence of overlapping, without modifying existing algorithms or data conformations as suggested by previous works. To test the suitability of the OVO approach with overlapping data, and due to the lack of proposals in the specialized literature, this research also introduces a novel scheme to artificially induce overlapping in real-world datasets, which enables us to simulate different types and levels of overlapping among the classes. The results obtained show that the methods using the OVO achieve better performances when considering data with overlapped classes than those dealing with all classes at the same time.
Open Access
Generación ilimitada de personajes mediante Stable Diffusion con DreamBooth y LoRA
(CAEPIA, 2024) Pascual Casas, Rubén; Maiza Coupin, Adrián Mikel; Sesma Sara, Mikel; Paternain Dallo, Daniel; Galar Idoate, Mikel; Estadística, Informática y Matemáticas; Estatistika, Informatika eta Matematika; Institute of Smart Cities - ISC; Universidad Pública de Navarra / Nafarroako Unibertsitate Publikoa, PJUPNA2023-11377; Gobierno de Navarra / Nafarroako Gobernua
Este artículo aborda el reto de generar un número ilimitado de personajes nuevos, y distintos, que engloben el estilo y las características visuales compartidas de un conjunto limitado de personajes diseñados por un humano. Este es un problema de gran relevancia en la industria audiovisual, ya que la capacidad de producir rápidamente personajes originales que se adhieran a unas características específicas aumenta enormemente las posibilidades en la producción de películas, series o videojuegos. Nuestra solución se basa en DreamBooth, un método de ajuste de modelos generativos de texto a imagen ampliamente extendido. Proponemos una adaptación centrada en dos retos principales: lo poco práctico que resulta utilizar prompts detallados de las imágenes para describir los personajes y la complejidad del ajuste de modelos a partir de un conjunto limitado de personajes. Para resolver estos problemas, introducimos en el entrenamiento de DreamBooth tokens adicionales específicos para cada personaje y eliminamos el conjunto de datos de regularización. Para generar personajes de manera ilimitada, proponemos el uso de tokens y embeddings aleatorios. Comprobamos la utilidad de la propuesta utilizando dos conjuntos de datos diferentes. Los resultados obtenidos muestran la capacidad de nuestro método para producir personajes diversos que se adhieren a un estilo y a unas características visuales concretas. Finalmente, desarrollamos un estudio de ablación.
Open Access
d-Choquet integrals: Choquet integrals based on dissimilarities
(Elsevier, 2020) Bustince Sola, Humberto; Mesiar, Radko; Fernández Fernández, Francisco Javier; Galar Idoate, Mikel; Paternain Dallo, Daniel; Altalhi, A. H.; Pereira Dimuro, Graçaliz; Bedregal, Benjamin; Takáč, Zdenko; Estatistika, Informatika eta Matematika; Institute of Smart Cities - ISC; Estadística, Informática y Matemáticas; Universidad Pública de Navarra / Nafarroako Unibertsitate Publikoa, PJUPNA13
The paper introduces a new class of functions from [0,1]n to [0,n] called d-Choquet integrals. These functions are a generalization of the 'standard' Choquet integral obtained by replacing the difference in the definition of the usual Choquet integral by a dissimilarity function. In particular, the class of all d-Choquet integrals encompasses the class of all 'standard' Choquet integrals but the use of dissimilarities provides higher flexibility and generality. We show that some d-Choquet integrals are aggregation/pre-aggregation/averaging/functions and some of them are not. The conditions under which this happens are stated and other properties of the d-Choquet integrals are studied.
Open Access
A study of OWA operators learned in convolutional neural networks
(MDPI, 2021) Domínguez Catena, Iris; Paternain Dallo, Daniel; Galar Idoate, Mikel; Institute of Smart Cities - ISC; Universidad Pública de Navarra / Nafarroako Unibertsitate Publikoa
Ordered Weighted Averaging (OWA) operators have been integrated in Convolutional Neural Networks (CNNs) for image classification through the OWA layer. This layer lets the CNN integrate global information about the image in the early stages, where most CNN architectures only allow for the exploitation of local information. As a side effect of this integration, the OWA layer becomes a practical method for the determination of OWA operator weights, which is usually a difficult task that complicates the integration of these operators in other fields. In this paper, we explore the weights learned for the OWA operators inside the OWA layer, characterizing them through their basic properties of orness and dispersion. We also compare them to some families of OWA operators, namely the Binomial OWA operator, the Stancu OWA operator and the expo-nential RIM OWA operator, finding examples that are currently impossible to generalize through these parameterizations.
Open Access
On the influence of interval normalization in IVOVO fuzzy multi-class classifier
(Springer, 2019) Uriz Martín, Mikel Xabier; Paternain Dallo, Daniel; Bustince Sola, Humberto; Galar Idoate, Mikel; Estatistika, Informatika eta Matematika; Institute of Smart Cities - ISC; Estadística, Informática y Matemáticas; Universidad Pública de Navarra / Nafarroako Unibertsitate Publikoa, PJUPNA13
IVOVO stands for Inverval-Valued One-Vs-One and is the combination of IVTURS fuzzy classifier and the One-Vs-One strategy. This method is designed to improve the performance of IVTURS in multi-class problems, by dividing the original problem into simpler binary ones. The key issue with IVTURS is that interval-valued confidence degrees for each class are returned and, consequently, they have to be normalized for applying a One-Vs-One strategy. However, there is no consensus on which normalization method should be used with intervals. In IVOVO, the normalization method based on the upper bounds was considered as it maintains the admissible order between intervals and also the proportion of ignorance, but no further study was developed. In this work, we aim to extend this analysis considering several normalizations in the literature. We will study both their main theoretical properties and empirical performance in the final results of IVOVO.
Open Access
Bitcoin and cybersecurity: temporal dissection of blockchain data to unveil changes in entity behavioral patterns
(MDPI, 2019) Zola, Francesco; Bruse, Jan Lukas; Eguimendia, María; Galar Idoate, Mikel; Orduna Urrutia, Raúl; Institute of Smart Cities - ISC
The Bitcoin network not only is vulnerable to cyber-attacks but currently represents the most frequently used cryptocurrency for concealing illicit activities. Typically, Bitcoin activity is monitored by decreasing anonymity of its entities using machine learning-based techniques, which consider the whole blockchain. This entails two issues: first, it increases the complexity of the analysis requiring higher efforts and, second, it may hide network micro-dynamics important for detecting short-term changes in entity behavioral patterns. The aim of this paper is to address both issues by performing a 'temporal dissection' of the Bitcoin blockchain, i.e., dividing it into smaller temporal batches to achieve entity classification. The idea is that a machine learning model trained on a certain time-interval (batch) should achieve good classification performance when tested on another batch if entity behavioral patterns are similar. We apply cascading machine learning principles'a type of ensemble learning applying stacking techniques'introducing a 'k-fold cross-testing' concept across batches of varying size. Results show that blockchain batch size used for entity classification could be reduced for certain classes (Exchange, Gambling, and eWallet) as classification rates did not vary significantly with batch size; suggesting that behavioral patterns did not change significantly over time. Mixer and Market class detection, however, can be negatively affected. A deeper analysis of Mining Pool behavior showed that models trained on recent data perform better than models trained on older data, suggesting that 'typical' Mining Pool behavior may be represented better by recent data. This work provides a first step towards uncovering entity behavioral changes via temporal dissection of blockchain data.