Jose Estevez

This paper systematically evaluates saliency methods as explainability tools for convolutional neural networks trained to diagnose glaucoma using simplified eye fundus images that contain only disc and cup outlines. These simplified images, a methodological novelty, were used to relate features highlighted in the saliency maps to the geometrical clues that experts consider in glaucoma diagnosis. Despite their simplicity, these images retained sufficient information for accurate classification, with balanced accuracies ranging from 0.8331 to 0.8890, compared to 0.8090 to 0.9203 for networks trained on the original images. The study used a dataset of 606 images, along with RIM-ONE DL and REFUGE datasets, and explored nine saliency methods. A discretization algorithm was applied to reduce noise and compute normalized attribution values for standard eye fundus sectors. Consistent with other medical imaging studies, significant variability was found in the attribution maps, influenced by the method, model, or architecture, and often deviating from typical sectors experts examine. However, globally, the results were relatively stable, with a strong correlation of 0.9289 (p < 0.001) between relevant sectors in our dataset and RIM-ONE DL, and 0.7806 (p < 0.001) for REFUGE. The findings suggest caution when using saliency methods in critical fields like medicine. These methods may be more suitable for broad image relevance interpretation rather than assessing individual cases, where results are highly sensitive to methodological choices. Moreover, the regions identified by the networks do not consistently align with established medical criteria for disease severity.

Deep learning systems, especially in critical fields like medicine, suffer from a significant drawback, their black box nature, which lacks mechanisms for explaining or interpreting their decisions. In this regard, our research aims to evaluate the use of surrogate models for interpreting convolutional neural network (CNN) decisions in glaucoma diagnosis. Our approach is novel in that we approximate the original model with an interpretable one and also change the input features, replacing pixels with tabular geometric features of the optic disc, cup, and neuroretinal rim. We trained CNNs with two types of images: original images of the optic nerve head and simplified images showing only the disc and cup contours on a uniform background. Decision trees were used as surrogate models due to their simplicity and visualization properties, while saliency maps were calculated for some images for comparison. The experiments carried out with 1271 images of healthy subjects and 721 images of glaucomatous eyes demonstrate that decision trees can closely approximate the predictions of neural networks trained on simplified contour images, with R-squared values near 0.9 for VGG19, Resnet50, InceptionV3 and Xception architectures. Saliency maps proved difficult to interpret and showed inconsistent results across architectures, in contrast to the decision trees. Additionally, some decision trees trained as surrogate models outperformed a decision tree trained on the actual outcomes without surrogation. Decision trees may be a more interpretable alternative to saliency methods. Moreover, the fact that we matched the performance of a decision tree without surrogation to that obtained by decision trees using knowledge distillation from neural networks is a great advantage since decision trees are inherently interpretable. Therefore, based on our findings, we think this approach would be the most recommendable choice for specialists as a diagnostic tool.

This paper describes a localization module for an autonomous wheelchair. This module includes a combination of various sensors such as odometers, laser scanners, IMU and Doppler speed sensors. Every sensor used in the module features variable covariance estimation in order to yield a final accurate localization. The main problem of a localization module composed of different sensors is the accuracy estimation of each sensor. Average static values are normally used, but these can lead to failure in some situations. In this paper, all the sensors have a variable covariance estimation that depends on the data quality. A Doppler speed sensor is used to estimate the covariance of the encoder odometric localization. Lidar is also used as a scan matching localization algorithm, comparing the difference between two consecutive scans to obtain the change in position. Matching quality gives the accuracy of the scan matcher localization. This structure yields a better position than a traditional odometric static covariance method. This is tested in a real prototype and compared to a standard fusion technique.

This paper researches a collaborative strategy between an XCS learning classifier system (LCS) and a relational learning (RL) agent. The problem here is to learn a relational policy for a stochastic markovian decision process. In the proposed method the XCS agent is used to improve the performance of the RL agent by filtering the samples used at the induction step. This research shows that in these conditions, one of the main benefits of using the XCS algorithm comes from selecting the examples for relational learning using an estimation for the accuracy of the predicted value at each state-action pair. This kind of transfer learning is important because the characteristics of both agents are complementary: the RL agent incrementally induces a high level description of a policy, while the LCS agent offers adaptation to changes in the environment.

Transfer learning, using systems with rich and general representations, to improve adaptive rule based systems designed to efficiently react in changing environments is the idea behind the problem studied in this paper. In this framework, the aim of this research is studying the benefits of using relational learning in combination with an evolutionary propositional learning system as XCS. The proposed method starts by learning a first order relational decission tree using a set of simplified instances of a problem. The learned relational model is then used to help a learning classifier system to deal with a more complex instance of the task. The researched strategy is based on injecting rules derived from the relational model in the discovering subsystem of the XCS. Results show that this method can be used to automatically adapt the behaviour of a learning rule based system when the environment increases its complexity.

The aim of this work is to provide a new approach to the classical problem of determining whether or not a set of data has been sampled from a univariate normal distribution. A simple neural network architecture is proposed as an efficient way of com

The aim of this paper is to provide a novel approach to the classical problem of testing the Gaussianity of a stationary process where the data can be assumed to be independent and identically distributed. Some results of Large Deviations theory are used to select a set of asymptotically optimal test statistics so that their combination results in a suitable detection procedure. The Johnson system of distributions has been used as the basis for the design of the detector. The fact that the good asymptotical properties can be extended to finite samples with small sizes provides the motivation for the rest of the work. In this respect, the performance of the proposed procedure is also compared with that of some known Gaussianity tests. The possibility to directly calculate a reliable measure of support such as the Bayesian posterior probabilities, as opposed to P-values, is an additional advantage. Furthermore, it will be shown how the proposed procedure can provide some extra information concerning the type of deviation from Gaussianity, which is present in the data.