Research Papers

Understanding how information is represented in neural networks is a fundamental challenge in both neuroscience and artificial intelligence. Despite their nonlinear architectures, recent evidence suggests that neural networks encode features in superposition, meaning that input concepts are linearly overlaid within the network's representations. We present a perspective that explains this phenomenon and provides a foundation for extracting interpretable representations from neural activations. Our theoretical framework consists of three steps: (1) Identifiability theory shows that neural networks trained for classification recover latent features up to a linear transformation. (2) Sparse coding methods can extract disentangled features from these representations by leveraging principles from compressed sensing. (3) Quantitative interpretability metrics provide a means to assess the success of these methods, ensuring that extracted features align with human-interpretable concepts. By bridging insights from theoretical neuroscience, representation learning, and interpretability research, we propose an emerging perspective on understanding neural representations in both artificial and biological systems. Our arguments have implications for neural coding theories, AI transparency, and the broader goal of making deep learning models more interpretable.

Explingo: Explaining AI Predictions using Large Language Models

Alexandra Zytek, Sara Pido, Sarah Alnegheimish +2

Explanations of machine learning (ML) model predictions generated by Explainable AI (XAI) techniques such as SHAP are essential for people using ML outputs for decision-making. We explore the potential of Large Language Models (LLMs) to transform these explanations into human-readable, narrative formats that align with natural communication. We address two key research questions: (1) Can LLMs reliably transform traditional explanations into high-quality narratives? and (2) How can we effectively evaluate the quality of narrative explanations? To answer these questions, we introduce Explingo, which consists of two LLM-based subsystems, a Narrator and Grader. The Narrator takes in ML explanations and transforms them into natural-language descriptions. The Grader scores these narratives on a set of metrics including accuracy, completeness, fluency, and conciseness. Our experiments demonstrate that LLMs can generate high-quality narratives that achieve high scores across all metrics, particularly when guided by a small number of human-labeled and bootstrapped examples. We also identified areas that remain challenging, in particular for effectively scoring narratives in complex domains. The findings from this work have been integrated into an open-source tool that makes narrative explanations available for further applications.

The goal of the present study is to analyze an application of deep learning models in order to augment the diagnostic performance of oral squamous cell carcinoma (OSCC) with a longitudinal cohort study using the Histopathological Imaging Database for oral cancer analysis. The dataset consisted of 5192 images (2435 Normal and 2511 OSCC), which were allocated between training, testing, and validation sets with an estimated ratio repartition of about 52% for the OSCC group, and still, our performance measure was validated on a combination set that contains almost equal number of sample in this use case as entire database have been divided into half using stratified splitting technique based again near binary proportion but total distribution was around even. We selected four deep-learning architectures for evaluation in the present study: ResNet101, DenseNet121, VGG16, and EfficientnetB3. EfficientNetB3 was found to be the best, with an accuracy of 98.33% and F1 score (0.9844), and it took remarkably less computing power in comparison with other models. The subsequent one was DenseNet121, with 90.24% accuracy and an F1 score of 90.45%. Moreover, we employed the Local Interpretable Model-agnostic Explanations (LIME) method to clarify why EfficientNetB3 made certain decisions with its predictions to improve the explainability and trustworthiness of results. This work provides evidence for the possible superior diagnosis in OSCC activated from the EfficientNetB3 model with the explanation of AI techniques such as LIME and paves an important groundwork to build on towards clinical usage.

We present a new technique that explains the output of a CNN-based model using a combination of GradCAM and LRP methods. Both of these methods produce visual explanations by highlighting input regions that are important for predictions. In the new method, the explanation produced by GradCAM is first processed to remove noises. The processed output is then multiplied elementwise with the output of LRP. Finally, a Gaussian blur is applied on the product. We compared the proposed method with GradCAM and LRP on the metrics of Faithfulness, Robustness, Complexity, Localisation and Randomisation. It was observed that this method performs better on Complexity than both GradCAM and LRP and is better than atleast one of them in the other metrics.

As Transformers have become state-of-the-art models for natural language processing (NLP) tasks, the need to understand and explain their predictions is increasingly apparent. Especially in unsupervised applications, such as information retrieval tasks, similarity models built on top of foundation model representations have been widely applied. However, their inner prediction mechanisms have mostly remained opaque. Recent advances in explainable AI have made it possible to mitigate these limitations by leveraging improved explanations for Transformers through layer-wise relevance propagation (LRP). Using BiLRP, an extension developed for computing second-order explanations in bilinear similarity models, we investigate which feature interactions drive similarity in NLP models. We validate the resulting explanations and demonstrate their utility in three corpus-level use cases, analyzing grammatical interactions, multilingual semantics, and biomedical text retrieval. Our findings contribute to a deeper understanding of different semantic similarity tasks and models, highlighting how novel explainable AI methods enable in-depth analyses and corpus-level insights.

Understanding Large Language Model Behaviors through Interactive Counterfactual Generation and Analysis

Furui Cheng, Vilém Zouhar, Robin Shing Moon Chan +3

Understanding the behavior of large language models (LLMs) is crucial for ensuring their safe and reliable use. However, existing explainable AI (XAI) methods for LLMs primarily rely on word-level explanations, which are often computationally inefficient and misaligned with human reasoning processes. Moreover, these methods often treat explanation as a one-time output, overlooking its inherently interactive and iterative nature. In this paper, we present LLM Analyzer, an interactive visualization system that addresses these limitations by enabling intuitive and efficient exploration of LLM behaviors through counterfactual analysis. Our system features a novel algorithm that generates fluent and semantically meaningful counterfactuals via targeted removal and replacement operations at user-defined levels of granularity. These counterfactuals are used to compute feature attribution scores, which are then integrated with concrete examples in a table-based visualization, supporting dynamic analysis of model behavior. A user study with LLM practitioners and interviews with experts demonstrate the system's usability and effectiveness, emphasizing the importance of involving humans in the explanation process as active participants rather than passive recipients.

This paper explores the integration of Explainable Automated Machine Learning (AutoML) in the realm of financial engineering, specifically focusing on its application in credit decision-making. The rapid evolution of Artificial Intelligence (AI) in finance has necessitated a balance between sophisticated algorithmic decision-making and the need for transparency in these systems. The focus is on how AutoML can streamline the development of robust machine learning models for credit scoring, while Explainable AI (XAI) methods, particularly SHapley Additive exPlanations (SHAP), provide insights into the models' decision-making processes. This study demonstrates how the combination of AutoML and XAI not only enhances the efficiency and accuracy of credit decisions but also fosters trust and collaboration between humans and AI systems. The findings underscore the potential of explainable AutoML in improving the transparency and accountability of AI-driven financial decisions, aligning with regulatory requirements and ethical considerations.

Shapley values, originating in game theory and increasingly prominent in explainable AI, have been proposed to assess the contribution of facts in query answering over databases, along with other similar power indices such as Banzhaf values. In this work we adapt these Shapley-like scores to probabilistic settings, the objective being to compute their expected value. We show that the computations of expected Shapley values and of the expected values of Boolean functions are interreducible in polynomial time, thus obtaining the same tractability landscape. We investigate the specific tractable case where Boolean functions are represented as deterministic decomposable circuits, designing a polynomial-time algorithm for this setting. We present applications to probabilistic databases through database provenance, and an effective implementation of this algorithm within the ProvSQL system, which experimentally validates its feasibility over a standard benchmark.

Background: Accurate survival time estimates aid end-of-life medical decision-making. Objectives: Develop an interpretable survival model for elderly residential aged care residents using advanced machine learning. Setting: A major Australasian residential aged care provider. Participants: Residents aged 65+ admitted for long-term care from July 2017 to August 2023. Sample size: 11,944 residents across 40 facilities. Predictors: Factors include age, gender, health status, co-morbidities, cognitive function, mood, nutrition, mobility, smoking, sleep, skin integrity, and continence. Outcome: Probability of survival post-admission, specifically calibrated for 6-month survival estimates. Statistical Analysis: Tested CoxPH, EN, RR, Lasso, GB, XGB, and RF models in 20 experiments with a 90/10 train/test split. Evaluated accuracy using C-index, Harrell's C-index, dynamic AUROC, IBS, and calibrated ROC. Chose XGB for its performance and calibrated it for 1, 3, 6, and 12-month predictions using Platt scaling. Employed SHAP values to analyze predictor impacts. Results: GB, XGB, and RF models showed the highest C-Index values (0.714, 0.712, 0.712). The optimal XGB model demonstrated a 6-month survival prediction AUROC of 0.746 (95% CI 0.744-0.749). Key mortality predictors include age, male gender, mobility, health status, pressure ulcer risk, and appetite. Conclusions: The study successfully applies machine learning to create a survival model for aged care, aligning with clinical insights on mortality risk factors and enhancing model interpretability and clinical utility through explainable AI.

Language models have the potential to assess mental health using social media data. By analyzing online posts and conversations, these models can detect patterns indicating mental health conditions like depression, anxiety, or suicidal thoughts. They examine keywords, language markers, and sentiment to gain insights into an individual's mental well-being. This information is crucial for early detection, intervention, and support, improving mental health care and prevention strategies. However, using language models for mental health assessments from social media has two limitations: (1) They do not compare posts against clinicians' diagnostic processes, and (2) It's challenging to explain language model outputs using concepts that the clinician can understand, i.e., clinician-friendly explanations. In this study, we introduce Process Knowledge-infused Learning (PK-iL), a new learning paradigm that layers clinical process knowledge structures on language model outputs, enabling clinician-friendly explanations of the underlying language model predictions. We rigorously test our methods on existing benchmark datasets, augmented with such clinical process knowledge, and release a new dataset for assessing suicidality. PK-iL performs competitively, achieving a 70% agreement with users, while other XAI methods only achieve 47% agreement (average inter-rater agreement of 0.72). Our evaluations demonstrate that PK-iL effectively explains model predictions to clinicians.

Disproving XAI Myths with Formal Methods -- Initial Results

Joao Marques-Silva

The advances in Machine Learning (ML) in recent years have been both impressive and far-reaching. However, the deployment of ML models is still impaired by a lack of trust in how the best-performing ML models make predictions. The issue of lack of trust is even more acute in the uses of ML models in high-risk or safety-critical domains. eXplainable artificial intelligence (XAI) is at the core of ongoing efforts for delivering trustworthy AI. Unfortunately, XAI is riddled with critical misconceptions, that foster distrust instead of building trust. This paper details some of the most visible misconceptions in XAI, and shows how formal methods have been used, both to disprove those misconceptions, but also to devise practically effective alternatives.

COCKATIEL: COntinuous Concept ranKed ATtribution with Interpretable ELements for explaining neural net classifiers on NLP tasks

Fanny Jourdan, Agustin Picard, Thomas Fel +3

Transformer architectures are complex and their use in NLP, while it has engendered many successes, makes their interpretability or explainability challenging. Recent debates have shown that attention maps and attribution methods are unreliable (Pruthi et al., 2019; Brunner et al., 2019). In this paper, we present some of their limitations and introduce COCKATIEL, which successfully addresses some of them. COCKATIEL is a novel, post-hoc, concept-based, model-agnostic XAI technique that generates meaningful explanations from the last layer of a neural net model trained on an NLP classification task by using Non-Negative Matrix Factorization (NMF) to discover the concepts the model leverages to make predictions and by exploiting a Sensitivity Analysis to estimate accurately the importance of each of these concepts for the model. It does so without compromising the accuracy of the underlying model or requiring a new one to be trained. We conduct experiments in single and multi-aspect sentiment analysis tasks and we show COCKATIEL's superior ability to discover concepts that align with humans' on Transformer models without any supervision, we objectively verify the faithfulness of its explanations through fidelity metrics, and we showcase its ability to provide meaningful explanations in two different datasets.