Deep learning has made great progress in supervised image classification. However, a persistent challenge in this field is the need for better explainability. This is crucial for building trust, troubleshooting, and ensuring regulatory compliance, ultimately leading to more responsible and effective AI applications. To tackle this challenge, we have developed a deep learning framework based on concepts, a recent method aimed at enhancing explainability, albeit with less emphasis on image classification performance. Our innovation lies in integrating a specific knowledge graph as a reasoning tool. This approach leverages the interdependence between concepts and classes to improve the performance of the concept-based model, achieving a balance between explainability and efficiency. Results from both synthetic and real data validate the effectiveness of our approach in balancing efficiency and explainability. Additionally, our model supports human test-time intervention to update its final prediction after incorporating new expert feedback. Our experiments show significant improvements in both classification and concept efficiencies.

This paper introduces a deep learning method for image classification that leverages knowledge formalised as a graph created from information represented by pairs attribute/value. The proposed method investigates a loss function that adaptively combines the classical cross-entropy commonly used in deep learning with a novel penalty function. The novel loss function is derived from the representation of nodes after embedding the knowledge graph and incorporates the proximity between class and image nodes. Its formulation enables the model to focus on identifying the boundary between the most challenging classes to distinguish. Experimental results on several image databases demonstrate improved performance compared to state-of-the-art methods, including classical deep learning algorithms and recent algorithms that incorporate knowledge represented by a graph.

Early diagnosis of osteoporosis can efficiently predict fracture risk. There is a great demand to prevent this disease. The goal of this study was to distinguish osteoporotic cases from healthy controls on 2D bone radiograph images, using texture analysis and genetic algorithms (GAs). Gray Level Co-occurrence Matrix (GLCM), Run length Matrix (RLM) and Binarized Statistical Image Features (BSIF) were used for texture analysis. Features are numerous and parameter-dependent. The related experts can pick out the useful input features for the classifier. It however remains a difficult task and may be inefficient or even harmful as the data pattern is not clear. In this paper, GAs were used to optimize the two parameters of the co-occurrence matrix (distance parameter or pixel separation, orientation or direction) and the number of gray levels used in the preprocessing quantification step. GAs were also used to select the best combination of features extracted from GLCM and RLM matrices. Experiments were conducted on two populations composed of Osteoporotic Patients and Control Subjects. Results show that GAs combined with GLCM and BSIF features can improve the classification rates (ACC = 87.50%) obtained using GLCM (ACC = 77.8%) alone.

New clustering algorithms are expected to manage complex data, meaning various shapes and densities while being user friendly. This work addresses this challenge. A new clustering algorithm KdMutual 1 driven by the number of clusters is proposed. The idea behind the algorithm is based on the assumption that working with cluster cores rather than considering frontiers makes the clustering process easier. KdMutual is based on three steps: The first one aims at identifying the potential core clusters. It relies on mutual neighborhood and includes specific mechanisms to identify and preserve potential core clusters. The second step is based on a constrained hierarchical process that deals with noise. In the last step the potential clusters are selected using a specific ranking criterion and the final partition is built. KdMutual combines the best characteristics of density peaks and connectivity-based approaches. It is capable of detecting the non-presence of natural clusters. Tests were carried out to compare the proposal with 14 other clustering algorithms. Using 2-dimensional benchmark datasets of various shapes and densities they showed that KdMutual was highly effective in matching a ground truth target. It also proved efficient in high dimensions when clusters are well separated. Moreover, it is able to identify clusters of various densities, partially overlapping and including a large amount of noise within spaces of moderate dimension.

In this paper, a new instance selection algorithm is proposed in the context of classification to manage non-trivial database sizes. The algorithm is hybrid and runs with only a few parameters that directly control the balance between the three objectives of classification, i.e. errors, storage req uirements and runtime. It comprises different mechanisms involving neighborhood and stratification algorithms that specifically speed up the runtime without significantly degrading efficiency. Instead of applying an IS (Instance Selection) algorithm to the whole database, IS is applied to strata deriving from the regions, each region representing a set of patterns selected from the original training set. The application of IS is conditioned by the purity of each region (i.e. the extent to which different categories of patterns are mixed in the region) and the stratification strategy is adapted to the region components. For each region, the number of delivered instances is firstly limited via the use of an iterative process that takes into account the boundary complexity, and secondly optimized by removing the superfluous ones. The sets of instances determined from all the regions are put together to provide an intermediate instance set that undergoes a dedicated filtering process to deliver the final set. Experiments performed with various synthetic and real data sets demonstrate the advantages of the proposed approach.

The design of finite impulse response (FIR) filters can be formulated as a non‐linear optimization problem reputed to be difficult for conventional approaches. The constraints are high and a large number of parameters have to be estimated, especially when dealing with two‐dimensional FIR filters. In order to improve the performance of conventional approaches, the authors explore several stochastic methodologies capable of handling large spaces. The authors specifically propose a new genetic algorithm (GA) in which some innovative concepts are introduced to improve the convergence and make its use easier for practitioners. The algorithm is globally improved by adapting the mutation and crossover and selection operators with the genetic advances. A dynamic ranking selection scheme is introduced to limit the promotion of extraordinary chromosomes. A refreshing mechanism is investigated to manage the trade‐off between diversity and elitism. The key point of the proposed approach stems from the capacity of the GA to adapt the genetic operators during the genetic life while remaining simple and easy to implement. Most of the parameters and operators are changed by the GA itself. From an initial calibration, the GA performs the design problem while calibrating and repeatedly re‐calibrating itself for solving it. The authors demonstrate on various cases of filter design a significant improvement in performance.

This paper is devoted to feature and instance selection managed by genetic algorithms (GA) in the context of supervised classification. We propose a GA encoded for selecting features in which each evaluated chromosome delivers a set of instances. The main aim is to optimize the processing time, which is particularly problematic when handling large databases. A key feature of our approach is the variable fitness evaluation based on scalability methodologies. Experimental results indicate that the preliminary version of the proposed algorithm can significantly reduce the computation time and is therefore applicable to high-dimensional data sets.