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.

Abstract While most continual learning methods focus on mitigating forgetting and improving accuracy, they often overlook the critical aspect of network calibration, despite its importance. Neural collapse, a phenomenon where last-layer features collapse to their class means, has demonstrated advantages in continual learning by reducing feature-classifier misalignment. Few works aim to improve the calibration of continual models for more reliable predictions. Our work goes a step further by proposing a novel method that not only enhances calibration but also improves performance by reducing overconfidence, mitigating forgetting, and increasing accuracy. We introduce Sphere-Adaptive Mixup (SAMix), an adaptive mixup strategy tailored for neural collapse-based methods. SAMix adapts the mixing process to the geometric properties of feature spaces under neural collapse, ensuring more robust regularization and alignment. Experiments show that SAMix significantly boosts performance, surpassing SOTA methods in continual learning while also improving model calibration. SAMix enhances both across-task accuracy and the broader reliability of predictions, making it a promising advancement for robust continual learning systems.

Contrastive learning has significantly improved representation quality, enhancing knowledge transfer across tasks in continual learning (CL). However, catastrophic forgetting remains a key challenge, as contrastive based methods primarily focus on “soft relationships” or “softness” between samples, which shift with changing data distributions and lead to representation overlap across tasks. Recently, the newly identified Neural Collapse phenomenon has shown promise in CL by focusing on ““““hard relationships” or “hardness” between samples and fixed proto-types. However, this approach overlooks “softness”, crucial for capturing intra-class variability, and this rigid focus can also pull old class representations toward current ones, increasing forgetting. Building on these insights, we propose Focal Neural Collapse Contrastive$(FNC^{2})$, a novel representation learning loss that effectively balances both soft and hard relationships. Additionally, we introduce the Hardness-Softness Distillation (HSD) loss to progressively preserve the knowledge gained from these relationships across tasks. Our method outperforms state-of-the-art approaches, particularly in minimizing memory reliance. Remarkably, even without the use of memory, our approach rivals rehearsal-based methods, offering a compelling solution for data privacy concerns.

Constrained clustering algorithms integrate user knowledge as constraints in the clustering process to guide it towards a desired outcome. When interacting with users, it is essential to quickly ask simple questions to identify informative constraints that will efficiently enhance an initial partition. We propose a new active query strategy for incremental clustering that translates user feedback into interpretable decision rules and identifies relevant points for queries using rule-based heuristics Experiments on benchmark datasets highlight the benefits of our new approach, making it suitable for real-world applications.

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.

This paper proposes a tentative and original survey of meeting points between Knowledge Representation and Reasoning (KRR) and Machine Learning (ML), two areas which have been developed quite separately in the last four decades. First, some common concerns are identified and discussed such as the types of representation used, the roles of knowledge and data, the lack or the excess of information, or the need for explanations and causal understanding. Then, the survey is organised in seven sections covering most of the territory where KRR and ML meet. We start with a section dealing with prototypical approaches from the literature on learning and reasoning: Inductive Logic Programming, Statistical Relational Learning, and Neurosymbolic AI, where ideas from rule-based reasoning are combined with ML. Then we focus on the use of various forms of background knowledge in learning, ranging from additional regularisation terms in loss functions, to the problem of aligning symbolic and vector space representations, or the use of knowledge graphs for learning. Then, the next section describes how KRR notions may benefit to learning tasks. For instance, constraints can be used as in declarative data mining for influencing the learned patterns; or semantic features are exploited in low-shot learning to compensate for the lack of data; or yet we can take advantage of analogies for learning purposes. Conversely, another section investigates how ML methods may serve KRR goals. For instance, one may learn special kinds of rules such as default rules, fuzzy rules or threshold rules, or special types of information such as constraints, or preferences. The section also covers formal concept analysis and rough sets-based methods. Yet another section reviews various interactions between Automated Reasoning and ML, such as the use of ML methods in SAT solving to make reasoning faster. Then a section deals with works related to model accountability, including explainability and interpretability, fairness and robustness. Finally, a section covers works on handling imperfect or incomplete data, including the problem of learning from uncertain or coarse data, the use of belief functions for regression, a revision-based view of the EM algorithm, the use of possibility theory in statistics, or the learning of imprecise models. This paper thus aims at a better mutual understanding of research in KRR and ML, and how they can cooperate. The paper is completed by an abundant bibliography.

Clustering is an important Machine Learning task, which aims at discovering the implicit structure of data. Applying a clustering algorithm is easy but since clustering is an unsupervised task, tuning it so that the results is appropriate to the expert expectations is much less obvious. To overcome this, expert knowledge can be integrated into a clustering process; this is generally formalized as constraints on the desired output, thus leading to constrained clustering. There are two lines of research for clustering: distance based clustering, where data are grouped into clusters according to their dissimilarity and conceptual clustering, where a cluster must be a concept that is a set of objects and a set of properties that describe them. This second approach relies on Formal Concept Analysis and benefits from advances in Pattern Mining. [69] has shown the interest of declarative approaches for pattern mining and has led to a new research direction for clustering that is interested in the use of declarative frameworks, such as Integer Linear Programming, Constraint Programming or SAT for clustering. This has several advantages: finding a global optimum, integrating different kinds of constraints, even complex ones in a clustering process and even combining conceptual and distance-based clustering. In this paper we present an inventory of constraints and a survey of declarative methods for constrained clustering.

Clustering is a well-known task in Data Mining that aims at grouping data instances according to their similarity. It is an exploratory and unsupervised task whose results depend on many parameters, often requiring the expert to iterate several times before satisfaction. Constrained clustering has been introduced for better modeling the expectations of the expert. Nevertheless constrained clustering is not yet sufficient since it usually requires the constraints to be given before the clustering process. In this paper we address a more general problem that aims at modeling the exploratory clustering process, through a sequence of clustering modifications where expert constraints are added on the fly. We present an incremental constrained clustering framework integrating active query strategies and a Constraint Programming model to fit the expert expectations while preserving the stability of the partition, so that the expert can understand the process and apprehend its impact. Our model supports instance and group-level constraints, which can be relaxed. Experiments on reference datasets and a case study related to the analysis of satellite image time series show the relevance of our framework.

In the context of semi-supervised learning and clustering ensemble methods, we introduce a novel strategy to consider not only pairwise constraints, but also triplet constraints. As far as we are aware, the latter have not been addressed in the literature of semi-supervised clustering ensembles. The strategy consists of a post-processing applied once a consensus partition has been built. Taking into account the fact that the clusters of the consensus partition are usually not spherical, in order to maintain their complex shapes, we first generate anchors, which are data points judged representative of the clusters, in such a way that every point in the cluster has an anchor close to it. These anchors are then used to create a data structure that we call an allocation matrix, which measures the assignment score of each point to each cluster in the consensus partition. Such a matrix is provided to an Integer Linear Programing (ILP) model to find the partition which is closest to the consensus partition, while satisfying the constraints. The experimental results show that our method, given an initial consensus partition and a set of constraints, allows to satisfy all the given constraints, modifying the initial partition, but without deteriorating considerably its quality.

In many settings just finding a good clustering is insufficient and an explanation of the clustering is required. If the features used to perform the clustering are interpretable then methods such as conceptual clustering can be used. However, in many applications this is not the case particularly for image, graph and other complex data. Here we explore the setting where a set of interpretable discrete tags for each instance is available. We formulate the descriptive clustering problem as a bi-objective optimization to simultaneously find compact clusters using the features and to describe them using the tags. We present our formulation in a declarative platform and show it can be integrated into a standard iterative algorithm to find all Pareto optimal solutions to the two objectives. Preliminary results demonstrate the utility of our approach on real data sets for images and electronic health care records and that it outperforms single objective and multi-view clustering baselines.

Consider the situation where your favorite clustering algorithm applied to a data set returns a good clustering but there are a few undesirable properties. One adhoc way to fix this is to re-run the clustering algorithm and hope to find a better variation. Instead, we propose to not run the algorithm again but minimally modify the existing clustering to remove the undesirable properties. We formulate the minimal clustering modification problem where we are given an initial clustering produced from any algorithm. The clustering is then modified to: i) remove the undesirable properties and ii) be minimally different to the given clustering. We show the underlying feasibility sub-problem can be intractable and demonstrate the flexibility of our constraint programming formulation. We empirically validate its usefulness through experiments on social network and medical imaging data sets.

Consider if you wish to cluster your ego network in Facebook so as to find several useful groups each of which you can invite to a different dinner party. You may require that each cluster must contain equal number of males and females, that the width of a cluster in terms of age is at most 10 and that each person in a cluster should have at least r other people with the same hobby. These are examples of cardinality, geometric and density requirements/constraints respectfully that can make the clustering useful for a given purpose. However existing formulations of constrained clustering were not designed to handle these constraints since they typically deal with low-level, instance-level constraints. We formulate a constraint programming (CP) languages formulation of clustering with these cluster-level styles of constraints which we call actionable clustering. Experimental results show the potential uses of this work to make clustering more actionable. We also show that these constraints can be used to improve the accuracy of semi-supervised clustering.

Minimum sum-of-squares clustering (MSSC) is a widely studied task and numerous approximate as well as a number of exact algorithms have been developed for it. Recently the interest of integrating prior knowledge in data mining has been shown, and much attention has gone into incorporating user constraints into clustering algorithms in a generic way.

Constrained clustering is an important task in Data Mining. In the last ten years, many works have been done to extend classical clustering algorithms to handle user-defined constraints, but restricted to handle one kind of user-constraints. In a previous work [1], we have proposed a declarative and generic framework, based on Constraint Programming, which enables to design a clustering task by specifying an optimization criterion and different kinds of user-constraints. One of the criteria is the within-cluster sum of dissimilarities, which is represented by a sum constraint and reified equality constraints V=Σ1≤i<;j≤n(G[i]==G[j])aij· A direct implementation using predefined constraints is not effective as the propagation of theses constraints is weak. In this paper, we consider this criterion as a global constraint and develop a filtering algorithm for it. This filtering helps to improve significantly the model performance. Experiments on classical databases show the interest of our approach.

BACKGROUND: Gene regulatory network inference remains a challenging problem in systems biology despite the numerous approaches that have been proposed. When substantial knowledge on a gene regulatory network is already available, supervised network inference is appropriate. Such a method builds a binary classifier able to assign a class (Regulation/No regulation) to an ordered pair of genes. Once learnt, the pairwise classifier can be used to predict new regulations. In this work, we explore the framework of Markov Logic Networks (MLN) that combine features of probabilistic graphical models with the expressivity of first-order logic rules. RESULTS: We propose to learn a Markov Logic network, e.g. a set of weighted rules that conclude on the predicate "regulates", starting from a known gene regulatory network involved in the switch proliferation/differentiation of keratinocyte cells, a set of experimental transcriptomic data and various descriptions of genes all encoded into first-order logic. As training data are unbalanced, we use asymmetric bagging to learn a set of MLNs. The prediction of a new regulation can then be obtained by averaging predictions of individual MLNs. As a side contribution, we propose three in silico tests to assess the performance of any pairwise classifier in various network inference tasks on real datasets. A first test consists of measuring the average performance on balanced edge prediction problem; a second one deals with the ability of the classifier, once enhanced by asymmetric bagging, to update a given network. Finally our main result concerns a third test that measures the ability of the method to predict regulations with a new set of genes. As expected, MLN, when provided with only numerical discretized gene expression data, does not perform as well as a pairwise SVM in terms of AUPR. However, when a more complete description of gene properties is provided by heterogeneous sources, MLN achieves the same performance as a black-box model such as a pairwise SVM while providing relevant insights on the predictions. CONCLUSIONS: The numerical studies show that MLN achieves very good predictive performance while opening the door to some interpretability of the decisions. Besides the ability to suggest new regulations, such an approach allows to cross-validate experimental data with existing knowledge.

In this paper, we present a heuristic-based algorithm to learn discriminative MLN structures automatically, directly from a training dataset. The algorithm heuristically transforms the relational dataset into boolean tables from which it builds candidate clauses for learning the final MLN. Comparisons to the state-of-the-art structure learning algorithms for MLNs in the three real-world domains show that the proposed algorithm outperforms them in terms of the conditional log likelihood (CLL), and the area under the precision-recall curve (AUC).

It is well known that modeling with constraints networks require a fair expertise. Thus tools able to automatically generate such networks have gained a major interest. The major contribution of this paper is to set a new framework based on Inductive Logic Programming able to build a constraint model from solutions and non-solutions of related problems. The model is expressed in a middle-level modeling language. On this particular relational learning problem, traditional top-down search methods fall into blind search and bottom-up search methods produce too expensive coverage tests. Recent works in Inductive Logic Programming about phase transition and crossing plateau shows that no general solution can face all these difficulties. In this context, we have designed an algorithm combining the major qualities of these two types of search techniques. We present experimental results on some benchmarks ranging from puzzles to scheduling problems.