Measuring the alignment between representations lets us understand similarities between the feature spaces of different models, such as Vision Transformers trained under diverse paradigms. However, traditional measures for representational alignment yield only scalar values that obscure how these spaces agree in terms of learned features. To address this, we combine alignment analysis with concept discovery, allowing a fine-grained breakdown of alignment into individual concepts. This approach reveals both universal concepts across models and each representation's internal concept structure. We introduce a new definition of concepts as non-linear manifolds, hypothesizing they better capture the geometry of the featurespace. A sanity check demonstrates the advantage of this manifold-based definition over linear baselines for concept-based alignment. Finally, our alignment analysis of four different ViTs shows that increased supervision tends to reduce semantic organization in learned representations.

Preference mechanisms, such as human preference, LLM-as-a-Judge (LaaJ), and reward models, are central to aligning and evaluating large language models (LLMs). Yet, the underlying concepts that drive these preferences remain poorly understood. In this work, we propose a fully automated method for generating local and global concept-based explanations of preferences across multiple domains. Our method utilizes an LLM to identify concepts that distinguish between chosen and rejected responses, and to represent them with concept-based vectors. To model the relationships between concepts and preferences, we propose a white-box Hierarchical Multi-Domain Regression model that captures both domain-general and domain-specific effects. To evaluate our method, we curate a dataset spanning eight challenging and diverse domains and explain twelve mechanisms. Our method achieves strong preference prediction performance, outperforming baselines while also being explainable. Additionally, we assess explanations in two application-driven settings. First, guiding LLM outputs with concepts from LaaJ explanations yields responses that those judges consistently prefer. Second, prompting LaaJs with concepts explaining humans improves their preference predictions. Together, our work establishes a new paradigm for explainability in the era of LLMs.

With the rapid growth usage of face recognition in people's daily life, face anti-spoofing becomes increasingly important to avoid malicious attacks. Recent face anti-spoofing models can reach a high classification accuracy on multiple datasets but these models can only tell people "this face is fake" while lacking the explanation to answer "why it is fake". Such a system undermines trustworthiness and causes user confusion, as it denies their requests without providing any explanations. In this paper, we incorporate XAI into face anti-spoofing and propose a new problem termed X-FAS (eXplainable Face Anti-Spoofing) empowering face anti-spoofing models to provide an explanation. We propose SPED (SPoofing Evidence Discovery), an X-FAS method which can discover spoof concepts and provide reliable explanations on the basis of discovered concepts. To evaluate the quality of X-FAS methods, we propose an X-FAS benchmark with annotated spoofing evidence by experts. We analyze SPED explanations on face anti-spoofing dataset and compare SPED quantitatively and qualitatively with previous XAI methods on proposed X-FAS benchmark. Experimental results demonstrate SPED's ability to generate reliable explanations.

Vision transformers (ViTs) can be trained using various learning paradigms, from fully supervised to self-supervised. Diverse training protocols often result in significantly different feature spaces, which are usually compared through alignment analysis. However, current alignment measures quantify this relationship in terms of a single scalar value, obscuring the distinctions between common and unique features in pairs of representations that share the same scalar alignment. We address this limitation by combining alignment analysis with concept discovery, which enables a breakdown of alignment into single concepts encoded in feature space. This fine-grained comparison reveals both universal and unique concepts across different representations, as well as the internal structure of concepts within each of them. Our methodological contributions address two key prerequisites for concept-based alignment: 1) For a description of the representation in terms of concepts that faithfully capture the geometry of the feature space, we define concepts as the most general structure they can possibly form - arbitrary manifolds, allowing hidden features to be described by their proximity to these manifolds. 2) To measure distances between concept proximity scores of two representations, we use a generalized Rand index and partition it for alignment between pairs of concepts. We confirm the superiority of our novel concept definition for alignment analysis over existing linear baselines in a sanity check. The concept-based alignment analysis of representations from four different ViTs reveals that increased supervision correlates with a reduction in the semantic structure of learned representations.

Human cognition can leverage fundamental conceptual knowledge, like geometric and kinematic ones, to appropriately perceive, comprehend and interact with novel objects. Motivated by this finding, we aim to endow machine intelligence with an analogous capability through performing at the conceptual level, in order to understand and then interact with articulated objects, especially for those in novel categories, which is challenging due to the intricate geometric structures and diverse joint types of articulated objects. To achieve this goal, we propose Analytic Ontology Template (AOT), a parameterized and differentiable program description of generalized conceptual ontologies. A baseline approach called AOTNet driven by AOTs is designed accordingly to equip intelligent agents with these generalized concepts, and then empower the agents to effectively discover the conceptual knowledge on the structure and affordance of articulated objects. The AOT-driven approach yields benefits in three key perspectives: i) enabling concept-level understanding of articulated objects without relying on any real training data, ii) providing analytic structure information, and iii) introducing rich affordance information indicating proper ways of interaction. We conduct exhaustive experiments and the results demonstrate the superiority of our approach in understanding and then interacting with articulated objects.

Explaining the decisions of Deep Neural Networks (DNNs) for medical images has become increasingly important. Existing attribution methods have difficulty explaining the meaning of pixels while existing concept-based methods are limited by additional annotations or specific model structures that are difficult to apply to ultrasound images. In this paper, we propose the Lesion Concept Explainer (LCE) framework, which combines attribution methods with concept-based methods. We introduce the Segment Anything Model (SAM), fine-tuned on a large number of medical images, for concept discovery to enable a meaningful explanation of ultrasound image DNNs. The proposed framework is evaluated in terms of both faithfulness and understandability. We point out deficiencies in the popular faithfulness evaluation metrics and propose a new evaluation metric. Our evaluation of public and private breast ultrasound datasets (BUSI and FG-US-B) shows that LCE performs well compared to commonly-used explainability methods. Finally, we also validate that LCE can consistently provide reliable explanations for more meaningful fine-grained diagnostic tasks in breast ultrasound.

Do vision-language models (VLMs) pre-trained to caption an image of a "durian" learn visual concepts such as "brown" (color) and "spiky" (texture) at the same time? We aim to answer this question as visual concepts learned "for free" would enable wide applications such as neuro-symbolic reasoning or human-interpretable object classification. We assume that the visual concepts, if captured by pre-trained VLMs, can be extracted by their vision-language interface with text-based concept prompts. We observe that recent works prompting VLMs with concepts often differ in their strategies to define and evaluate the visual concepts, leading to conflicting conclusions. We propose a new concept definition strategy based on two observations: First, certain concept prompts include shortcuts that recognize correct concepts for wrong reasons; Second, multimodal information (e.g. visual discriminativeness, and textual knowledge) should be leveraged when selecting the concepts. Our proposed concept discovery and learning (CDL) framework is thus designed to identify a diverse list of generic visual concepts (e.g. "spiky" as opposed to "spiky durian"), which are ranked and selected based on visual and language mutual information. We carefully design quantitative and human evaluations of the discovered concepts on six diverse visual recognition datasets, which confirm that pre-trained VLMs do learn visual concepts that provide accurate and thorough descriptions for the recognized objects. All code and models are publicly released.

Trustworthiness is a major prerequisite for the safe application of opaque deep learning models in high-stakes domains like medicine. Understanding the decision-making process not only contributes to fostering trust but might also reveal previously unknown decision criteria of complex models that could advance the state of medical research. The discovery of decision-relevant concepts from black box models is a particularly challenging task. This study proposes Concept Discovery through Latent Diffusion-based Counterfactual Trajectories (CDCT), a novel three-step framework for concept discovery leveraging the superior image synthesis capabilities of diffusion models. In the first step, CDCT uses a Latent Diffusion Model (LDM) to generate a counterfactual trajectory dataset. This dataset is used to derive a disentangled representation of classification-relevant concepts using a Variational Autoencoder (VAE). Finally, a search algorithm is applied to identify relevant concepts in the disentangled latent space. The application of CDCT to a classifier trained on the largest public skin lesion dataset revealed not only the presence of several biases but also meaningful biomarkers. Moreover, the counterfactuals generated within CDCT show better FID scores than those produced by a previously established state-of-the-art method, while being 12 times more resource-efficient. Unsupervised concept discovery holds great potential for the application of trustworthy AI and the further development of human knowledge in various domains. CDCT represents a further step in this direction.

Adversarial attacks (AAs) pose a significant threat to the reliability and robustness of deep neural networks. While the impact of these attacks on model predictions has been extensively studied, their effect on the learned representations and concepts within these models remains largely unexplored. In this work, we perform an in-depth analysis of the influence of AAs on the concepts learned by convolutional neural networks (CNNs) using eXplainable artificial intelligence (XAI) techniques. Through an extensive set of experiments across various network architectures and targeted AA techniques, we unveil several key findings. First, AAs induce substantial alterations in the concept composition within the feature space, introducing new concepts or modifying existing ones. Second, the adversarial perturbation itself can be linearly decomposed into a set of latent vector components, with a subset of these being responsible for the attack's success. Notably, we discover that these components are target-specific, i.e., are similar for a given target class throughout different AA techniques and starting classes. Our findings provide valuable insights into the nature of AAs and their impact on learned representations, paving the way for the development of more robust and interpretable deep learning models, as well as effective defenses against adversarial threats.

Mara Graziani and colleagues Laura O' Mahony, An-Phi Nguyen, Henning Müller and Vincent Andrearczyk are researching deep learning models and the associated learned representations. In this interview, Mara tells us about the team's proposed framework for concept discovery. Our paper Uncovering Unique Concept Vectors through Latent Space Decomposition focuses on understanding how representations are organized by intermediate layers of complex deep learning models. The latent space of a layer can be interpreted as a vector space spanned by individual neuron directions. In our work, we identify a new basis that aligns with the variance of the training data.