Disentangled Explanations of Neural Network Predictions by Finding Relevant Subspaces.

Explainable AI aims to overcome the black-box nature of complex ML models like neural networks by generating explanations for their predictions. Explanations often take the form of a heatmap identifying input features (e.g. pixels) that are relevant to the model's decision. These explanations, however, entangle the potentially multiple factors that enter into the overall complex decision strategy. We propose to disentangle explanations by extracting at some intermediate layer of a neural network, subspaces that capture the multiple and distinct activation patterns (e.g. visual concepts) that are relevant to the prediction. To automatically extract these subspaces, we propose two new analyses, extending principles found in PCA or ICA to explanations. These novel analyses, which we call principal relevant component analysis (PRCA) and disentangled relevant subspace analysis (DRSA), maximize relevance instead of e.g. variance or kurtosis. This allows for a much stronger focus of the analysis on what the ML model actually uses for predicting, ignoring activations or concepts to which the model is invariant. Our approach is general enough to work alongside common attribution techniques such as Shapley Value, Integrated Gradients, or LRP. Our proposed methods show to be practically useful and compare favorably to the state of the art as demonstrated on benchmarks and three use cases.

[Explainable artificial intelligence in pathology]. / Erklärbare Künstliche Intelligenz in der Pathologie.

With the advancements in precision medicine, the demands on pathological diagnostics have increased, requiring standardized, quantitative, and integrated assessments of histomorphological and molecular pathological data. Great hopes are placed in artificial intelligence (AI) methods, which have demonstrated the ability to analyze complex clinical, histological, and molecular data for disease classification, biomarker quantification, and prognosis estimation. This paper provides an overview of the latest developments in pathology AI, discusses the limitations, particularly concerning the black box character of AI, and describes solutions to make decision processes more transparent using methods of so-called explainable AI (XAI).

Toward Explainable Artificial Intelligence for Precision Pathology.

The rapid development of precision medicine in recent years has started to challenge diagnostic pathology with respect to its ability to analyze histological images and increasingly large molecular profiling data in a quantitative, integrative, and standardized way. Artificial intelligence (AI) and, more precisely, deep learning technologies have recently demonstrated the potential to facilitate complex data analysis tasks, including clinical, histological, and molecular data for disease classification; tissue biomarker quantification; and clinical outcome prediction. This review provides a general introduction to AI and describes recent developments with a focus on applications in diagnostic pathology and beyond. We explain limitations including the black-box character of conventional AI and describe solutions to make machine learning decisions more transparent with so-called explainable AI. The purpose of the review is to foster a mutual understanding of both the biomedical and the AI side. To that end, in addition to providing an overview of the relevant foundations in pathology and machine learning, we present worked-through examples for a better practical understanding of what AI can achieve and how it should be done.

Learning domain invariant representations by joint Wasserstein distance minimization.

Domain shifts in the training data are common in practical applications of machine learning; they occur for instance when the data is coming from different sources. Ideally, a ML model should work well independently of these shifts, for example, by learning a domain-invariant representation. However, common ML losses do not give strong guarantees on how consistently the ML model performs for different domains, in particular, whether the model performs well on a domain at the expense of its performance on another domain. In this paper, we build new theoretical foundations for this problem, by contributing a set of mathematical relations between classical losses for supervised ML and the Wasserstein distance in joint space (i.e. representation and output space). We show that classification or regression losses, when combined with a GAN-type discriminator between domains, form an upper-bound to the true Wasserstein distance between domains. This implies a more invariant representation and also more stable prediction performance across domains. Theoretical results are corroborated empirically on several image datasets. Our proposed approach systematically produces the highest minimum classification accuracy across domains, and the most invariant representation.

Single-cell gene regulatory network prediction by explainable AI.

The molecular heterogeneity of cancer cells contributes to the often partial response to targeted therapies and relapse of disease due to the escape of resistant cell populations. While single-cell sequencing has started to improve our understanding of this heterogeneity, it offers a mostly descriptive view on cellular types and states. To obtain more functional insights, we propose scGeneRAI, an explainable deep learning approach that uses layer-wise relevance propagation (LRP) to infer gene regulatory networks from static single-cell RNA sequencing data for individual cells. We benchmark our method with synthetic data and apply it to single-cell RNA sequencing data of a cohort of human lung cancers. From the predicted single-cell networks our approach reveals characteristic network patterns for tumor cells and normal epithelial cells and identifies subnetworks that are observed only in (subgroups of) tumor cells of certain patients. While current state-of-the-art methods are limited by their ability to only predict average networks for cell populations, our approach facilitates the reconstruction of networks down to the level of single cells which can be utilized to characterize the heterogeneity of gene regulation within and across tumors.

From Clustering to Cluster Explanations via Neural Networks.

A recent trend in machine learning has been to enrich learned models with the ability to explain their own predictions. The emerging field of explainable AI (XAI) has so far mainly focused on supervised learning, in particular, deep neural network classifiers. In many practical problems, however, the label information is not given and the goal is instead to discover the underlying structure of the data, for example, its clusters. While powerful methods exist for extracting the cluster structure in data, they typically do not answer the question why a certain data point has been assigned to a given cluster. We propose a new framework that can, for the first time, explain cluster assignments in terms of input features in an efficient and reliable manner. It is based on the novel insight that clustering models can be rewritten as neural networks-or "neuralized." Cluster predictions of the obtained networks can then be quickly and accurately attributed to the input features. Several showcases demonstrate the ability of our method to assess the quality of learned clusters and to extract novel insights from the analyzed data and representations.

Patient-level proteomic network prediction by explainable artificial intelligence.

Understanding the pathological properties of dysregulated protein networks in individual patients' tumors is the basis for precision therapy. Functional experiments are commonly used, but cover only parts of the oncogenic signaling networks, whereas methods that reconstruct networks from omics data usually only predict average network features across tumors. Here, we show that the explainable AI method layer-wise relevance propagation (LRP) can infer protein interaction networks for individual patients from proteomic profiling data. LRP reconstructs average and individual interaction networks with an AUC of 0.99 and 0.93, respectively, and outperforms state-of-the-art network prediction methods for individual tumors. Using data from The Cancer Proteome Atlas, we identify known and potentially novel oncogenic network features, among which some are cancer-type specific and show only minor variation among patients, while others are present across certain tumor types but differ among individual patients. Our approach may therefore support predictive diagnostics in precision oncology by inferring "patient-level" oncogenic mechanisms.

Higher-Order Explanations of Graph Neural Networks via Relevant Walks.

Graph Neural Networks (GNNs) are a popular approach for predicting graph structured data. As GNNs tightly entangle the input graph into the neural network structure, common explainable AI approaches are not applicable. To a large extent, GNNs have remained black-boxes for the user so far. In this paper, we show that GNNs can in fact be naturally explained using higher-order expansions, i.e., by identifying groups of edges that jointly contribute to the prediction. Practically, we find that such explanations can be extracted using a nested attribution scheme, where existing techniques such as layer-wise relevance propagation (LRP) can be applied at each step. The output is a collection of walks into the input graph that are relevant for the prediction. Our novel explanation method, which we denote by GNN-LRP, is applicable to a broad range of graph neural networks and lets us extract practically relevant insights on sentiment analysis of text data, structure-property relationships in quantum chemistry, and image classification.

Building and Interpreting Deep Similarity Models.

Many learning algorithms such as kernel machines, nearest neighbors, clustering, or anomaly detection, are based on distances or similarities. Before similarities are used for training an actual machine learning model, we would like to verify that they are bound to meaningful patterns in the data. In this paper, we propose to make similarities interpretable by augmenting them with an explanation. We develop BiLRP, a scalable and theoretically founded method to systematically decompose the output of an already trained deep similarity model on pairs of input features. Our method can be expressed as a composition of LRP explanations, which were shown in previous works to scale to highly nonlinear models. Through an extensive set of experiments, we demonstrate that BiLRP robustly explains complex similarity models, e.g., built on VGG-16 deep neural network features. Additionally, we apply our method to an open problem in digital humanities: detailed assessment of similarity between historical documents, such as astronomical tables. Here again, BiLRP provides insight and brings verifiability into a highly engineered and problem-specific similarity model.

Machine learning analysis of DNA methylation profiles distinguishes primary lung squamous cell carcinomas from head and neck metastases.

Head and neck squamous cell carcinoma (HNSC) patients are at risk of suffering from both pulmonary metastases or a second squamous cell carcinoma of the lung (LUSC). Differentiating pulmonary metastases from primary lung cancers is of high clinical importance, but not possible in most cases with current diagnostics. To address this, we performed DNA methylation profiling of primary tumors and trained three different machine learning methods to distinguish metastatic HNSC from primary LUSC. We developed an artificial neural network that correctly classified 96.4% of the cases in a validation cohort of 279 patients with HNSC and LUSC as well as normal lung controls, outperforming support vector machines (95.7%) and random forests (87.8%). Prediction accuracies of more than 99% were achieved for 92.1% (neural network), 90% (support vector machine), and 43% (random forest) of these cases by applying thresholds to the resulting probability scores and excluding samples with low confidence. As independent clinical validation of the approach, we analyzed a series of 51 patients with a history of HNSC and a second lung tumor, demonstrating the correct classifications based on clinicopathological properties. In summary, our approach may facilitate the reliable diagnostic differentiation of pulmonary metastases of HNSC from primary LUSC to guide therapeutic decisions.

Unmasking Clever Hans predictors and assessing what machines really learn.

Current learning machines have successfully solved hard application problems, reaching high accuracy and displaying seemingly intelligent behavior. Here we apply recent techniques for explaining decisions of state-of-the-art learning machines and analyze various tasks from computer vision and arcade games. This showcases a spectrum of problem-solving behaviors ranging from naive and short-sighted, to well-informed and strategic. We observe that standard performance evaluation metrics can be oblivious to distinguishing these diverse problem solving behaviors. Furthermore, we propose our semi-automated Spectral Relevance Analysis that provides a practically effective way of characterizing and validating the behavior of nonlinear learning machines. This helps to assess whether a learned model indeed delivers reliably for the problem that it was conceived for. Furthermore, our work intends to add a voice of caution to the ongoing excitement about machine intelligence and pledges to evaluate and judge some of these recent successes in a more nuanced manner.

Computational analysis reveals histotype-dependent molecular profile and actionable mutation effects across cancers.

BACKGROUND: Comprehensive mutational profiling data now available on all major cancers have led to proposals of novel molecular tumor classifications that modify or replace the established organ- and tissue-based tumor typing. The rationale behind such molecular reclassifications is that genetic alterations underlying cancer pathology predict response to therapy and may therefore offer a more precise view on cancer than histology. The use of individual actionable mutations to select cancers for treatment across histotypes is already being tested in the so-called basket trials with variable success rates. Here, we present a computational approach that facilitates the systematic analysis of the histological context dependency of mutational effects by integrating genomic and proteomic tumor profiles across cancers. METHODS: To determine effects of oncogenic mutations on protein profiles, we used the energy distance, which compares the Euclidean distances of protein profiles in tumors with an oncogenic mutation (inner distance) to that in tumors without the mutation (outer distance) and performed Monte Carlo simulations for the significance analysis. Finally, the proteins were ranked by their contribution to profile differences to identify proteins characteristic of oncogenic mutation effects across cancers. RESULTS: We apply our approach to four current proposals of molecular tumor classifications and major therapeutically relevant actionable genes. All 12 actionable genes evaluated show effects on the protein level in the corresponding tumor type and showed additional mutation-related protein profiles in 21 tumor types. Moreover, our analysis identifies consistent cross-cancer effects for 4 genes (FGFR1, ERRB2, IDH1, KRAS/NRAS) in 14 tumor types. We further use cell line drug response data to validate our findings. CONCLUSIONS: This computational approach can be used to identify mutational signatures that have protein-level effects and can therefore contribute to preclinical in silico tests of the efficacy of molecular classifications as well as the druggability of individual mutations. It thus supports the identification of novel targeted therapies effective across cancers and guides efficient basket trial designs.

"What is relevant in a text document?": An interpretable machine learning approach.

Text documents can be described by a number of abstract concepts such as semantic category, writing style, or sentiment. Machine learning (ML) models have been trained to automatically map documents to these abstract concepts, allowing to annotate very large text collections, more than could be processed by a human in a lifetime. Besides predicting the text's category very accurately, it is also highly desirable to understand how and why the categorization process takes place. In this paper, we demonstrate that such understanding can be achieved by tracing the classification decision back to individual words using layer-wise relevance propagation (LRP), a recently developed technique for explaining predictions of complex non-linear classifiers. We train two word-based ML models, a convolutional neural network (CNN) and a bag-of-words SVM classifier, on a topic categorization task and adapt the LRP method to decompose the predictions of these models onto words. Resulting scores indicate how much individual words contribute to the overall classification decision. This enables one to distill relevant information from text documents without an explicit semantic information extraction step. We further use the word-wise relevance scores for generating novel vector-based document representations which capture semantic information. Based on these document vectors, we introduce a measure of model explanatory power and show that, although the SVM and CNN models perform similarly in terms of classification accuracy, the latter exhibits a higher level of explainability which makes it more comprehensible for humans and potentially more useful for other applications.

Evaluating the Visualization of What a Deep Neural Network Has Learned.

Deep neural networks (DNNs) have demonstrated impressive performance in complex machine learning tasks such as image classification or speech recognition. However, due to their multilayer nonlinear structure, they are not transparent, i.e., it is hard to grasp what makes them arrive at a particular classification or recognition decision, given a new unseen data sample. Recently, several approaches have been proposed enabling one to understand and interpret the reasoning embodied in a DNN for a single test image. These methods quantify the "importance" of individual pixels with respect to the classification decision and allow a visualization in terms of a heatmap in pixel/input space. While the usefulness of heatmaps can be judged subjectively by a human, an objective quality measure is missing. In this paper, we present a general methodology based on region perturbation for evaluating ordered collections of pixels such as heatmaps. We compare heatmaps computed by three different methods on the SUN397, ILSVRC2012, and MIT Places data sets. Our main result is that the recently proposed layer-wise relevance propagation algorithm qualitatively and quantitatively provides a better explanation of what made a DNN arrive at a particular classification decision than the sensitivity-based approach or the deconvolution method. We provide theoretical arguments to explain this result and discuss its practical implications. Finally, we investigate the use of heatmaps for unsupervised assessment of the neural network performance.Deep neural networks (DNNs) have demonstrated impressive performance in complex machine learning tasks such as image classification or speech recognition. However, due to their multilayer nonlinear structure, they are not transparent, i.e., it is hard to grasp what makes them arrive at a particular classification or recognition decision, given a new unseen data sample. Recently, several approaches have been proposed enabling one to understand and interpret the reasoning embodied in a DNN for a single test image. These methods quantify the "importance" of individual pixels with respect to the classification decision and allow a visualization in terms of a heatmap in pixel/input space. While the usefulness of heatmaps can be judged subjectively by a human, an objective quality measure is missing. In this paper, we present a general methodology based on region perturbation for evaluating ordered collections of pixels such as heatmaps. We compare heatmaps computed by three different methods on the SUN397, ILSVRC2012, and MIT Places data sets. Our main result is that the recently proposed layer-wise relevance propagation algorithm qualitatively and quantitatively provides a better explanation of what made a DNN arrive at a particular classification decision than the sensitivity-based approach or the deconvolution method. We provide theoretical arguments to explain this result and discuss its practical implications. Finally, we investigate the use of heatmaps for unsupervised assessment of the neural network performance.

On Pixel-Wise Explanations for Non-Linear Classifier Decisions by Layer-Wise Relevance Propagation.

Understanding and interpreting classification decisions of automated image classification systems is of high value in many applications, as it allows to verify the reasoning of the system and provides additional information to the human expert. Although machine learning methods are solving very successfully a plethora of tasks, they have in most cases the disadvantage of acting as a black box, not providing any information about what made them arrive at a particular decision. This work proposes a general solution to the problem of understanding classification decisions by pixel-wise decomposition of nonlinear classifiers. We introduce a methodology that allows to visualize the contributions of single pixels to predictions for kernel-based classifiers over Bag of Words features and for multilayered neural networks. These pixel contributions can be visualized as heatmaps and are provided to a human expert who can intuitively not only verify the validity of the classification decision, but also focus further analysis on regions of potential interest. We evaluate our method for classifiers trained on PASCAL VOC 2009 images, synthetic image data containing geometric shapes, the MNIST handwritten digits data set and for the pre-trained ImageNet model available as part of the Caffe open source package.

Assessment and Validation of Machine Learning Methods for Predicting Molecular Atomization Energies.

The accurate and reliable prediction of properties of molecules typically requires computationally intensive quantum-chemical calculations. Recently, machine learning techniques applied to ab initio calculations have been proposed as an efficient approach for describing the energies of molecules in their given ground-state structure throughout chemical compound space (Rupp et al. Phys. Rev. Lett. 2012, 108, 058301). In this paper we outline a number of established machine learning techniques and investigate the influence of the molecular representation on the methods performance. The best methods achieve prediction errors of 3 kcal/mol for the atomization energies of a wide variety of molecules. Rationales for this performance improvement are given together with pitfalls and challenges when applying machine learning approaches to the prediction of quantum-mechanical observables.