Research Papers

MERA: Multimodal and Multiscale Self-Explanatory Model with Considerably Reduced Annotation for Lung Nodule Diagnosis

Jiahao Lu, Chong Yin, Silvia Ingala +3

Lung cancer, a leading cause of cancer-related deaths globally, emphasises the importance of early detection for better patient outcomes. Pulmonary nodules, often early indicators of lung cancer, necessitate accurate, timely diagnosis. Despite Explainable Artificial Intelligence (XAI) advances, many existing systems struggle providing clear, comprehensive explanations, especially with limited labelled data. This study introduces MERA, a Multimodal and Multiscale self-Explanatory model designed for lung nodule diagnosis with considerably Reduced Annotation requirements. MERA integrates unsupervised and weakly supervised learning strategies (self-supervised learning techniques and Vision Transformer architecture for unsupervised feature extraction) and a hierarchical prediction mechanism leveraging sparse annotations via semi-supervised active learning in the learned latent space. MERA explains its decisions on multiple levels: model-level global explanations via semantic latent space clustering, instance-level case-based explanations showing similar instances, local visual explanations via attention maps, and concept explanations using critical nodule attributes. Evaluations on the public LIDC dataset show MERA's superior diagnostic accuracy and self-explainability. With only 1% annotated samples, MERA achieves diagnostic accuracy comparable to or exceeding state-of-the-art methods requiring full annotation. The model's inherent design delivers comprehensive, robust, multilevel explanations aligned closely with clinical practice, enhancing trustworthiness and transparency. Demonstrated viability of unsupervised and weakly supervised learning lowers the barrier to deploying diagnostic AI in broader medical domains. Our complete code is open-source available: https://github.com/diku-dk/credanno.

Explainability and its emerging counterpart contestability have become important normative and design principles for trustworthy AI as they enable users and subjects to understand and challenge AI decisions. However, realizing these principles is difficult, as they assume different meanings in technical, legal, and organizational dimensions of AI regulation. To resolve this conceptual polysemy, in this paper, we present the findings of an interview study with 14 experts to examine the intersection and implementation of explainability and contestability, and their understanding in different research communities. We outline differentiations between descriptive and normative explainability, judicial and non-judicial channels of contestation, and individual and collective contestation action. We further describe the main points of friction in the realization of both principles, including the alignment between top-down and bottom-up regulation, the assignment of responsibility, and the need for interdisciplinary collaboration. Lastly, we formulate three recommendations for AI policy to implement both principles through a Regulation by Design perspective. We believe our contributions can inform policy-making and regulation of these core principles and enable more effective and equitable design, development, and deployment of trustworthy public AI systems.

Explainable artificial intelligence (XAI) enhances AI system transparency by framing interpretability as an optimization problem. However, this approach often necessitates numerous iterations of computationally intensive operations, limiting its applicability in real-time scenarios. While recent research has focused on XAI hardware acceleration on FPGAs and TPU, these methods do not fully address energy efficiency in real-time settings. To address this limitation, we propose XAIedge, a novel framework that leverages approximate computing techniques into XAI algorithms, including integrated gradients, model distillation, and Shapley analysis. XAIedge translates these algorithms into approximate matrix computations and exploits the synergy between convolution, Fourier transform, and approximate computing paradigms. This approach enables efficient hardware acceleration on TPU-based edge devices, facilitating faster real-time outcome interpretations. Our comprehensive evaluation demonstrates that XAIedge achieves a $2\times$ improvement in energy efficiency compared to existing accurate XAI hardware acceleration techniques while maintaining comparable accuracy. These results highlight the potential of XAIedge to significantly advance the deployment of explainable AI in energy-constrained real-time applications.

Rule-based systems are a very popular form of explainable AI, particularly in the fuzzy community, where fuzzy rules are widely used for control and classification problems. However, fuzzy rule-based classifiers struggle to reach bigger traction outside of fuzzy venues, because users sometimes do not know about fuzzy and because fuzzy partitions are not so easy to interpret in some situations. In this work, we propose a methodology to reduce fuzzy rule-based classifiers to crisp rule-based classifiers. We study different possible crisp descriptions and implement an algorithm to obtain them. Also, we analyze the complexity of the resulting crisp classifiers. We believe that our results can help both fuzzy and non-fuzzy practitioners understand better the way in which fuzzy rule bases partition the feature space and how easily one system can be translated to another and vice versa. Our complexity metric can also help to choose between different fuzzy classifiers based on what the equivalent crisp partitions look like.

This work aligns deep learning (DL) with human reasoning capabilities and needs to enable more efficient, interpretable, and robust image classification. We approach this from three perspectives: explainability, causality, and biological vision. Introduction and background open this work before diving into operative chapters. First, we assess neural networks' visualization techniques for medical images and validate an explainable-by-design method for breast mass classification. A comprehensive review at the intersection of XAI and causality follows, where we introduce a general scaffold to organize past and future research, laying the groundwork for our second perspective. In the causality direction, we propose novel modules that exploit feature co-occurrence in medical images, leading to more effective and explainable predictions. We further introduce CROCODILE, a general framework that integrates causal concepts, contrastive learning, feature disentanglement, and prior knowledge to enhance generalization. Lastly, we explore biological vision, examining how humans recognize objects, and propose CoCoReco, a connectivity-inspired network with context-aware attention mechanisms. Overall, our key findings include: (i) simple activation maximization lacks insight for medical imaging DL models; (ii) prototypical-part learning is effective and radiologically aligned; (iii) XAI and causal ML are deeply connected; (iv) weak causal signals can be leveraged without a priori information to improve performance and interpretability; (v) our framework generalizes across medical domains and out-of-distribution data; (vi) incorporating biological circuit motifs improves human-aligned recognition. This work contributes toward human-aligned DL and highlights pathways to bridge the gap between research and clinical adoption, with implications for improved trust, diagnostic accuracy, and safe deployment.

Current AI systems based on probabilistic neural networks, such as large language models (LLMs), have demonstrated remarkable generative capabilities yet face critical challenges including hallucination, unpredictability, and misalignment with human decision-making. These issues fundamentally stem from the over-reliance on randomized (probabilistic) neural networks-oversimplified models of biological neural networks-while neglecting the role of procedural reasoning (chain-of-thought) in trustworthy decision-making. Inspired by the human cognitive duality of fluid intelligence (flexible generation) and crystallized intelligence (structured knowledge), this study proposes a dual-channel intelligent architecture that integrates probabilistic generation (LLMs) with white-box procedural reasoning (chain-of-thought) to construct interpretable, continuously learnable, and human-aligned AI systems. Concretely, this work: (1) redefines chain-of-thought as a programmable crystallized intelligence carrier, enabling dynamic knowledge evolution and decision verification through multi-turn interaction frameworks; (2) introduces a task-driven modular network design that explicitly demarcates the functional boundaries between randomized generation and procedural control to address trustworthiness in vertical-domain applications; (3) demonstrates that multi-turn interaction is a necessary condition for intelligence emergence, with dialogue depth positively correlating with the system's human-alignment degree. This research not only establishes a new paradigm for trustworthy AI deployment but also provides theoretical foundations for next-generation human-AI collaborative systems.

Proofs as Explanations: Short Certificates for Reliable Predictions

Avrim Blum, Steve Hanneke, Chirag Pabbaraju +1

We consider a model for explainable AI in which an explanation for a prediction $h(x)=y$ consists of a subset $S'$ of the training data (if it exists) such that all classifiers $h' \in H$ that make at most $b$ mistakes on $S'$ predict $h'(x)=y$. Such a set $S'$ serves as a proof that $x$ indeed has label $y$ under the assumption that (1) the target function $h^\star$ belongs to $H$, and (2) the set $S$ contains at most $b$ corrupted points. For example, if $b=0$ and $H$ is the family of linear classifiers in $\mathbb{R}^d$, and if $x$ lies inside the convex hull of the positive data points in $S$ (and hence every consistent linear classifier labels $x$ as positive), then Carathéodory's theorem states that $x$ lies inside the convex hull of $d+1$ of those points. So, a set $S'$ of size $d+1$ could be released as an explanation for a positive prediction, and would serve as a short proof of correctness of the prediction under the assumption of realizability. In this work, we consider this problem more generally, for general hypothesis classes $H$ and general values $b\geq 0$. We define the notion of the robust hollow star number of $H$ (which generalizes the standard hollow star number), and show that it precisely characterizes the worst-case size of the smallest certificate achievable, and analyze its size for natural classes. We also consider worst-case distributional bounds on certificate size, as well as distribution-dependent bounds that we show tightly control the sample size needed to get a certificate for any given test example. In particular, we define a notion of the certificate coefficient $\varepsilon_x$ of an example $x$ with respect to a data distribution $D$ and target function $h^\star$, and prove matching upper and lower bounds on sample size as a function of $\varepsilon_x$, $b$, and the VC dimension $d$ of $H$.

Uncovering the Structure of Explanation Quality with Spectral Analysis

Johannes Maeß, Grégoire Montavon, Shinichi Nakajima +2

As machine learning models are increasingly considered for high-stakes domains, effective explanation methods are crucial to ensure that their prediction strategies are transparent to the user. Over the years, numerous metrics have been proposed to assess quality of explanations. However, their practical applicability remains unclear, in particular due to a limited understanding of which specific aspects each metric rewards. In this paper we propose a new framework based on spectral analysis of explanation outcomes to systematically capture the multifaceted properties of different explanation techniques. Our analysis uncovers two distinct factors of explanation quality-stability and target sensitivity-that can be directly observed through spectral decomposition. Experiments on both MNIST and ImageNet show that popular evaluation techniques (e.g., pixel-flipping, entropy) partially capture the trade-offs between these factors. Overall, our framework provides a foundational basis for understanding explanation quality, guiding the development of more reliable techniques for evaluating explanations.

On Background Bias of Post-Hoc Concept Embeddings in Computer Vision DNNs

Gesina Schwalbe, Georgii Mikriukov, Edgar Heinert +5

The thriving research field of concept-based explainable artificial intelligence (C-XAI) investigates how human-interpretable semantic concepts embed in the latent spaces of deep neural networks (DNNs). Post-hoc approaches therein use a set of examples to specify a concept, and determine its embeddings in DNN latent space using data driven techniques. This proved useful to uncover biases between different target (foreground or concept) classes. However, given that the background is mostly uncontrolled during training, an important question has been left unattended so far: Are/to what extent are state-of-the-art, data-driven post-hoc C-XAI approaches themselves prone to biases with respect to their backgrounds? E.g., wild animals mostly occur against vegetation backgrounds, and they seldom appear on roads. Even simple and robust C-XAI methods might abuse this shortcut for enhanced performance. A dangerous performance degradation of the concept-corner cases of animals on the road could thus remain undiscovered. This work validates and thoroughly confirms that established Net2Vec-based concept segmentation techniques frequently capture background biases, including alarming ones, such as underperformance on road scenes. For the analysis, we compare 3 established techniques from the domain of background randomization on >50 concepts from 2 datasets, and 7 diverse DNN architectures. Our results indicate that even low-cost setups can provide both valuable insight and improved background robustness.

Despite the impressive performance of large language models (LLMs), they can present unintended biases and harmful behaviors driven by encoded values, emphasizing the urgent need to understand the value mechanisms behind them. However, current research primarily evaluates these values through external responses with a focus on AI safety, lacking interpretability and failing to assess social values in real-world contexts. In this paper, we propose a novel framework called ValueExploration, which aims to explore the behavior-driven mechanisms of National Social Values within LLMs at the neuron level. As a case study, we focus on Chinese Social Values and first construct C-voice, a large-scale bilingual benchmark for identifying and evaluating Chinese Social Values in LLMs. By leveraging C-voice, we then identify and locate the neurons responsible for encoding these values according to activation difference. Finally, by deactivating these neurons, we analyze shifts in model behavior, uncovering the internal mechanism by which values influence LLM decision-making. Extensive experiments on four representative LLMs validate the efficacy of our framework. The benchmark and code will be available.

Explaining Low Perception Model Competency with High-Competency Counterfactuals

Sara Pohland, Claire Tomlin

There exist many methods to explain how an image classification model generates its decision, but very little work has explored methods to explain why a classifier might lack confidence in its prediction. As there are various reasons the classifier might lose confidence, it would be valuable for this model to not only indicate its level of uncertainty but also explain why it is uncertain. Counterfactual images have been used to visualize changes that could be made to an image to generate a different classification decision. In this work, we explore the use of counterfactuals to offer an explanation for low model competency--a generalized form of predictive uncertainty that measures confidence. Toward this end, we develop five novel methods to generate high-competency counterfactual images, namely Image Gradient Descent (IGD), Feature Gradient Descent (FGD), Autoencoder Reconstruction (Reco), Latent Gradient Descent (LGD), and Latent Nearest Neighbors (LNN). We evaluate these methods across two unique datasets containing images with six known causes for low model competency and find Reco, LGD, and LNN to be the most promising methods for counterfactual generation. We further evaluate how these three methods can be utilized by pre-trained Multimodal Large Language Models (MLLMs) to generate language explanations for low model competency. We find that the inclusion of a counterfactual image in the language model query greatly increases the ability of the model to generate an accurate explanation for the cause of low model competency, thus demonstrating the utility of counterfactual images in explaining low perception model competency.

Multimodal prediction models based on imaging and clinical data are increasingly used for clinical decision support, yet their interpretability remains limited. We present multimodal Deep Transformation Models (DTMs) combining statistical approaches and neural networks to achieve strong predictive performance while preserving interpretability for tabular data. A key contribution of this work is the adaption of the xAI methods Grad-CAM and Occlusion to DTMs relying on 3D CNNs, enabling interpretation of the image branch through the generation of explanation maps. We developed DTMs to predict functional independence three months after stroke using diffusion-weighted imaging and clinical data from 407 patients. In a ten-fold cross-validation, the models achieved state-of-the-art predictive performance (AUC 0.81 [0.75, 0.87]) while maintaining interpretability for tabular features, with functional independence before stroke and stroke severity on admission emerging as the strongest predictors. Explanation maps from both xAI methods highlighted consistent regions, including frontal lobe areas which are known to be associated with age, a strong predictor of functional outcome. Notably, these regions disappeared once age was included as an explicit tabular predictor. Similarity analyses of explanation maps revealed distinct spatial patterns, providing meaningful insights into stroke pathophysiology, systematic error analysis and hypothesis generation.