Machine Learning (ML) models are increasingly used to make critical decisions in real-world applications, yet they have become more complex, making them harder to understand. To this end, researchers have proposed several techniques to explain model predictions. However, practitioners struggle to use these explainability techniques because they often do not know which one to choose and how to interpret the results of the explanations. In this work, we address these challenges by introducing TalkToModel: an interactive dialogue system for explaining machine learning models through conversations. Specifically, TalkToModel comprises of three key components: 1) a natural language interface for engaging in conversations, making ML model explainability highly accessible, 2) a dialogue engine that adapts to any tabular model and dataset, interprets natural language, maps it to appropriate explanations, and generates text responses, and 3) an execution component that constructs the explanations. We carried out extensive quantitative and human subject evaluations of TalkToModel. Overall, we found the conversational system understands user inputs on novel datasets and models with high accuracy, demonstrating the system's capacity to generalize to new situations. In real-world evaluations with humans, 73% of healthcare workers (e.g., doctors and nurses) agreed they would use TalkToModel over baseline point-and-click systems for explainability in a disease prediction task, and 85% of ML professionals agreed TalkToModel was easier to use for computing explanations. Our findings demonstrate that TalkToModel is more effective for model explainability than existing systems, introducing a new category of explainability tools for practitioners. Code & demo released here: https://github.com/dylan-slack/TalkToModel.
Both in the domains of Feature Selection and Interpretable AI, there exists a desire to `rank' features based on their importance. Such feature importance rankings can then be used to either: (1) reduce the dataset size or (2) interpret the Machine Learning model. In the literature, however, such Feature Rankers are not evaluated in a systematic, consistent way. Many papers have a different way of arguing which feature importance ranker works best. This paper fills this gap, by proposing a new evaluation methodology. By making use of synthetic datasets, feature importance scores can be known beforehand, allowing more systematic evaluation. To facilitate large-scale experimentation using the new methodology, a benchmarking framework was built in Python, called fseval. The framework allows running experiments in parallel and distributed over machines on HPC systems. By integrating with an online platform called Weights and Biases, charts can be interactively explored on a live dashboard. The software was released as open-source software, and is published as a package on the PyPi platform. The research concludes by exploring one such large-scale experiment, to find the strengths and weaknesses of the participating algorithms, on many fronts.
Explanations for \emph{black-box} models help us understand model decisions as well as provide information on model biases and inconsistencies. Most of the current explainability techniques provide a single level of explanation, often in terms of feature importance scores or feature attention maps in input space. Our focus is on explaining deep discriminative models at \emph{multiple levels of abstraction}, from fine-grained to fully abstract explanations. We achieve this by using the natural properties of \emph{hyperbolic geometry} to more efficiently model a hierarchy of symbolic features and generate \emph{hierarchical symbolic rules} as part of our explanations. Specifically, for any given deep discriminative model, we distill the underpinning knowledge by discretisation of the continuous latent space using vector quantisation to form symbols, followed by a \emph{hyperbolic reasoning block} to induce an \emph{abstraction tree}. We traverse the tree to extract explanations in terms of symbolic rules and its corresponding visual semantics. We demonstrate the effectiveness of our method on the MNIST and AFHQ high-resolution animal faces dataset. Our framework is available at \url{https://github.com/koriavinash1/SymbolicInterpretability}.
Complex machine learning algorithms are used more and more often in critical tasks involving text data, leading to the development of interpretability methods. Among local methods, two families have emerged: those computing importance scores for each feature and those extracting simple logical rules. In this paper we show that using different methods can lead to unexpectedly different explanations, even when applied to simple models for which we would expect qualitative coincidence. To quantify this effect, we propose a new approach to compare explanations produced by different methods.
Neural networks are ubiquitous in applied machine learning for education. Their pervasive success in predictive performance comes alongside a severe weakness, the lack of explainability of their decisions, especially relevant in human-centric fields. We implement five state-of-the-art methodologies for explaining black-box machine learning models (LIME, PermutationSHAP, KernelSHAP, DiCE, CEM) and examine the strengths of each approach on the downstream task of student performance prediction for five massive open online courses. Our experiments demonstrate that the families of explainers do not agree with each other on feature importance for the same Bidirectional LSTM models with the same representative set of students. We use Principal Component Analysis, Jensen-Shannon distance, and Spearman's rank-order correlation to quantitatively cross-examine explanations across methods and courses. Furthermore, we validate explainer performance across curriculum-based prerequisite relationships. Our results come to the concerning conclusion that the choice of explainer is an important decision and is in fact paramount to the interpretation of the predictive results, even more so than the course the model is trained on. Source code and models are released at http://github.com/epfl-ml4ed/evaluating-explainers.
Supervised learning can be viewed as distilling relevant information from input data into feature representations. This process becomes difficult when supervision is noisy as the distilled information might not be relevant. In fact, recent research shows that networks can easily overfit all labels including those that are corrupted, and hence can hardly generalize to clean datasets. In this paper, we focus on the problem of learning with noisy labels and introduce compression inductive bias to network architectures to alleviate this over-fitting problem. More precisely, we revisit one classical regularization named Dropout and its variant Nested Dropout. Dropout can serve as a compression constraint for its feature dropping mechanism, while Nested Dropout further learns ordered feature representations w.r.t. feature importance. Moreover, the trained models with compression regularization are further combined with Co-teaching for performance boost. Theoretically, we conduct bias-variance decomposition of the objective function under compression regularization. We analyze it for both single model and Co-teaching. This decomposition provides three insights: (i) it shows that over-fitting is indeed an issue for learning with noisy labels; (ii) through an information bottleneck formulation, it explains why the proposed feature compression helps in combating label noise; (iii) it gives explanations on the performance boost brought by incorporating compression regularization into Co-teaching. Experiments show that our simple approach can have comparable or even better performance than the state-of-the-art methods on benchmarks with real-world label noise including Clothing1M and ANIMAL-10N. Our implementation is available at https://yingyichen-cyy.github.io/CompressFeatNoisyLabels/.
Machine learning (ML) offers powerful methods for detecting and modeling associations often in data with large feature spaces and complex associations. Many useful tools/packages (e.g. scikit-learn) have been developed to make the various elements of data handling, processing, modeling, and interpretation accessible. However, it is not trivial for most investigators to assemble these elements into a rigorous, replicatable, unbiased, and effective data analysis pipeline. Automated machine learning (AutoML) seeks to address these issues by simplifying the process of ML analysis for all. Here, we introduce STREAMLINE, a simple, transparent, end-to-end AutoML pipeline designed as a framework to easily conduct rigorous ML modeling and analysis (limited initially to binary classification). STREAMLINE is specifically designed to compare performance between datasets, ML algorithms, and other AutoML tools. It is unique among other autoML tools by offering a fully transparent and consistent baseline of comparison using a carefully designed series of pipeline elements including: (1) exploratory analysis, (2) basic data cleaning, (3) cross validation partitioning, (4) data scaling and imputation, (5) filter-based feature importance estimation, (6) collective feature selection, (7) ML modeling with `Optuna' hyperparameter optimization across 15 established algorithms (including less well-known Genetic Programming and rule-based ML), (8) evaluation across 16 classification metrics, (9) model feature importance estimation, (10) statistical significance comparisons, and (11) automatically exporting all results, plots, a PDF summary report, and models that can be easily applied to replication data.
Many past works aim to improve visual reasoning in models by supervising feature importance (estimated by model explanation techniques) with human annotations such as highlights of important image regions. However, recent work has shown that performance gains from feature importance (FI) supervision for Visual Question Answering (VQA) tasks persist even with random supervision, suggesting that these methods do not meaningfully align model FI with human FI. In this paper, we show that model FI supervision can meaningfully improve VQA model accuracy as well as performance on several Right-for-the-Right-Reason (RRR) metrics by optimizing for four key model objectives: (1) accurate predictions given limited but sufficient information (Sufficiency); (2) max-entropy predictions given no important information (Uncertainty); (3) invariance of predictions to changes in unimportant features (Invariance); and (4) alignment between model FI explanations and human FI explanations (Plausibility). Our best performing method, Visual Feature Importance Supervision (VisFIS), outperforms strong baselines on benchmark VQA datasets in terms of both in-distribution and out-of-distribution accuracy. While past work suggests that the mechanism for improved accuracy is through improved explanation plausibility, we show that this relationship depends crucially on explanation faithfulness (whether explanations truly represent the model's internal reasoning). Predictions are more accurate when explanations are plausible and faithful, and not when they are plausible but not faithful. Lastly, we show that, surprisingly, RRR metrics are not predictive of out-of-distribution model accuracy when controlling for a model's in-distribution accuracy, which calls into question the value of these metrics for evaluating model reasoning. All supporting code is available at https://github.com/zfying/visfis
Although semantic communications have exhibited satisfactory performance for a large number of tasks, the impact of semantic noise and the robustness of the systems have not been well investigated. Semantic noise refers to the misleading between the intended semantic symbols and received ones, thus cause the failure of tasks. In this paper, we first propose a framework for the robust end-to-end semantic communication systems to combat the semantic noise. In particular, we analyze sample-dependent and sample-independent semantic noise. To combat the semantic noise, the adversarial training with weight perturbation is developed to incorporate the samples with semantic noise in the training dataset. Then, we propose to mask a portion of the input, where the semantic noise appears frequently, and design the masked vector quantized-variational autoencoder (VQ-VAE) with the noise-related masking strategy. We use a discrete codebook shared by the transmitter and the receiver for encoded feature representation. To further improve the system robustness, we develop a feature importance module (FIM) to suppress the noise-related and task-unrelated features. Thus, the transmitter simply needs to transmit the indices of these important task-related features in the codebook. Simulation results show that the proposed method can be applied in many downstream tasks and significantly improve the robustness against semantic noise with remarkable reduction on the transmission overhead.
Decoding brain imaging data are gaining popularity, with applications in brain-computer interfaces and the study of neural representations. Decoding is typicallysubject-specific and does not generalise well over subjects, due to high amounts ofbetween subject variability. Techniques that overcome this will not only providericher neuroscientific insights but also make it possible for group-level models to out-perform subject-specific models. Here, we propose a method that uses subjectembedding, analogous to word embedding in natural language processing, to learnand exploit the structure in between-subject variability as part of a decoding model,our adaptation of the WaveNet architecture for classification. We apply this to mag-netoencephalography data, where 15 subjects viewed 118 different images, with30 examples per image; to classify images using the entire 1 s window followingimage presentation. We show that the combination of deep learning and subjectembedding is crucial to closing the performance gap between subject- and group-level decoding models. Importantly, group models outperform subject models onlow-accuracy subjects (although slightly impair high-accuracy subjects) and can behelpful for initialising subject models. While we have not generally found group-levelmodels to perform better than subject-level models, the performance of groupmodelling is expected to be even higher with bigger datasets. In order to providephysiological interpretation at the group level, we make use of permutation featureimportance. This provides insights into the spatiotemporal and spectral informationencoded in the models. All code is available on GitHub (https://github.com/ricsinaruto/MEG-group-decode).
Not all real-world data are labeled, and when labels are not available, it is often costly to obtain them. Moreover, as many algorithms suffer from the curse of dimensionality, reducing the features in the data to a smaller set is often of great utility. Unsupervised feature selection aims to reduce the number of features, often using feature importance scores to quantify the relevancy of single features to the task at hand. These scores can be based only on the distribution of variables and the quantification of their interactions. The previous literature, mainly investigating anomaly detection and clusters, fails to address the redundancy-elimination issue. We propose an evaluation of correlations among features to compute feature importance scores representing the contribution of single features in explaining the dataset's structure. Based on Coalitional Game Theory, our feature importance scores include a notion of redundancy awareness making them a tool to achieve redundancy-free feature selection. We show that the deriving features' selection outperforms competing methods in lowering the redundancy rate while maximizing the information contained in the data. We also introduce an approximated version of the algorithm to reduce the complexity of Shapley values' computations.
Predicting the duration of traffic incidents is a challenging task due to the stochastic nature of events. The ability to accurately predict how long accidents will last can provide significant benefits to both end-users in their route choice and traffic operation managers in handling of non-recurrent traffic congestion. This paper presents a novel bi-level machine learning framework enhanced with outlier removal and intra-extra joint optimisation for predicting the incident duration on three heterogeneous data sets collected for both arterial roads and motorways from Sydney, Australia and San-Francisco, U.S.A. Firstly, we use incident data logs to develop a binary classification prediction approach, which allows us to classify traffic incidents as short-term or long-term. We find the optimal threshold between short-term versus long-term traffic incident duration, targeting both class balance and prediction performance while also comparing the binary versus multi-class classification approaches. Secondly, for more granularity of the incident duration prediction to the minute level, we propose a new Intra-Extra Joint Optimisation algorithm (IEO-ML) which extends multiple baseline ML models tested against several regression scenarios across the data sets. Final results indicate that: a) 40-45 min is the best split threshold for identifying short versus long-term incidents and that these incidents should be modelled separately, b) our proposed IEO-ML approach significantly outperforms baseline ML models in $66\%$ of all cases showcasing its great potential for accurate incident duration prediction. Lastly, we evaluate the feature importance and show that time, location, incident type, incident reporting source and weather at among the top 10 critical factors which influence how long incidents will last.
The analysis of electrocardiogram (ECG) signals can be time consuming as it is performed manually by cardiologists. Therefore, automation through machine learning (ML) classification is being increasingly proposed which would allow ML models to learn the features of a heartbeat and detect abnormalities. The lack of interpretability hinders the application of Deep Learning in healthcare. Through interpretability of these models, we would understand how a machine learning algorithm makes its decisions and what patterns are being followed for classification. This thesis builds Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) classifiers based on state-of-the-art models and compares their performance and interpretability to shallow classifiers. Here, both global and local interpretability methods are exploited to understand the interaction between dependent and independent variables across the entire dataset and to examine model decisions in each sample, respectively. Partial Dependence Plots, Shapley Additive Explanations, Permutation Feature Importance, and Gradient Weighted Class Activation Maps (Grad-Cam) are the four interpretability techniques implemented on time-series ML models classifying ECG rhythms. In particular, we exploit Grad-Cam, which is a local interpretability technique and examine whether its interpretability varies between correctly and incorrectly classified ECG beats within each class. Furthermore, the classifiers are evaluated using K-Fold cross-validation and Leave Groups Out techniques, and we use non-parametric statistical testing to examine whether differences are significant. It was found that Grad-CAM was the most effective interpretability technique at explaining predictions of proposed CNN and LSTM models. We concluded that all high performing classifiers looked at the QRS complex of the ECG rhythm when making predictions.
Surrogate models are data-based approximations of computationally expensive simulations that enable efficient exploration of the model's design space and informed decision-making in many physical domains. The usage of surrogate models in the vibroacoustic domain, however, is challenging due to the non-smooth, complex behavior of wave phenomena. This paper investigates four Machine Learning (ML) approaches in the modelling of surrogates of Sound Transmission Loss (STL). Feature importance and feature engineering are used to improve the models' accuracy while increasing their interpretability and physical consistency. The transfer of the proposed techniques to other problems in the vibroacoustic domain and possible limitations of the models are discussed.
Scientists frequently prioritize learning from data rather than training the best possible model; however, research in machine learning often prioritizes the latter. Marginal contribution feature importance (MCI) was developed to break this trend by providing a useful framework for quantifying the relationships in data. In this work, we aim to improve upon the theoretical properties, performance, and runtime of MCI by introducing ultra-marginal feature importance (UMFI), which uses dependence removal techniques from the AI fairness literature as its foundation. We first propose axioms for feature importance methods that seek to explain the causal and associative relationships in data, and we prove that UMFI satisfies these axioms under basic assumptions. We then show on real and simulated data that UMFI performs better than MCI, especially in the presence of correlated interactions and unrelated features, while partially learning the structure of the causal graph and reducing the exponential runtime of MCI to super-linear.
In this paper, we investigated whether we can 1) detect participants with ataxia-specific gait characteristics (risk-prediction), and 2) assess severity of ataxia from gait (severity-assessment) using computer vision. We created a dataset of 155 videos from 89 participants, 24 controls and 65 diagnosed with (or are pre-manifest) spinocerebellar ataxias (SCAs), performing the gait task of the Scale for the Assessment and Rating of Ataxia (SARA) from 11 medical sites located in 8 different states across the United States. We develop a computer vision pipeline to detect, track, and separate out the participants from their surroundings and construct several features from their body pose coordinates to capture gait characteristics like step width, step length, swing, stability, speed, etc. Our risk-prediction model achieves 83.06% accuracy and an 80.23% F1 score. Similarly, our severity-assessment model achieves a mean absolute error (MAE) score of 0.6225 and a Pearson's correlation coefficient score of 0.7268. Our models still performed competitively when evaluated on data from sites not used during training. Furthermore, through feature importance analysis, we found that our models associate wider steps, decreased walking speed, and increased instability with greater ataxia severity, which is consistent with previously established clinical knowledge. Our models create possibilities for remote ataxia assessment in non-clinical settings in the future, which could significantly improve accessibility of ataxia care. Furthermore, our underlying dataset was assembled from a geographically diverse cohort, highlighting its potential to further increase equity. The code used in this study is open to the public, and the anonymized body pose landmark dataset is also available upon request.
Being able to explain the prediction to clinical end-users is a necessity to leverage the power of artificial intelligence (AI) models for clinical decision support. For medical images, a feature attribution map, or heatmap, is the most common form of explanation that highlights important features for AI models' prediction. However, it is unknown how well heatmaps perform on explaining decisions on multi-modal medical images, where each image modality or channel visualizes distinct clinical information of the same underlying biomedical phenomenon. Understanding such modality-dependent features is essential for clinical users' interpretation of AI decisions. To tackle this clinically important but technically ignored problem, we propose the modality-specific feature importance (MSFI) metric. It encodes clinical image and explanation interpretation patterns of modality prioritization and modality-specific feature localization. We conduct a clinical requirement-grounded, systematic evaluation using computational methods and a clinician user study. Results show that the examined 16 heatmap algorithms failed to fulfill clinical requirements to correctly indicate AI model decision process or decision quality. The evaluation and MSFI metric can guide the design and selection of XAI algorithms to meet clinical requirements on multi-modal explanation.
In this work, we propose two novel methodologies to study temporal and morphological phenotypic effects caused by different experimental conditions using imaging data. As a proof of concept, we apply them to analyze drug effects in 2D cancer cell cultures. We train a convolutional autoencoder on 1M images dataset with random augmentations and multi-crops to use as feature extractor. We systematically compare it to the pretrained state-of-the-art models. We further use the feature extractor in two ways. First, we apply distance-based analysis and dynamic time warping to cluster temporal patterns of 31 drugs. We identify clusters allowing annotation of drugs as having cytotoxic, cytostatic, mixed or no effect. Second, we implement an adversarial/regularized learning setup to improve classification of 31 drugs and visualize image regions that contribute to the improvement. We increase top-3 classification accuracy by 8% on average and mine examples of morphological feature importance maps. We provide the feature extractor and the weights to foster transfer learning applications in biology. We also discuss utility of other pretrained models and applicability of our methods to other types of biomedical data.
Unsupervised black-box models are challenging to interpret. Indeed, most existing explainability methods require labels to select which component(s) of the black-box's output to interpret. In the absence of labels, black-box outputs often are representation vectors whose components do not correspond to any meaningful quantity. Hence, choosing which component(s) to interpret in a label-free unsupervised/self-supervised setting is an important, yet unsolved problem. To bridge this gap in the literature, we introduce two crucial extensions of post-hoc explanation techniques: (1) label-free feature importance and (2) label-free example importance that respectively highlight influential features and training examples for a black-box to construct representations at inference time. We demonstrate that our extensions can be successfully implemented as simple wrappers around many existing feature and example importance methods. We illustrate the utility of our label-free explainability paradigm through a qualitative and quantitative comparison of representation spaces learned by various autoencoders trained on distinct unsupervised tasks.
Although media content is increasingly produced, distributed, and consumed in multiple combinations of modalities, how individual modalities contribute to the perceived emotion of a media item remains poorly understood. In this paper we present MusicVideos (MuVi), a novel dataset for affective multimedia content analysis to study how the auditory and visual modalities contribute to the perceived emotion of media. The data were collected by presenting music videos to participants in three conditions: music, visual, and audiovisual. Participants annotated the music videos for valence and arousal over time, as well as the overall emotion conveyed. We present detailed descriptive statistics for key measures in the dataset and the results of feature importance analyses for each condition. Finally, we propose a novel transfer learning architecture to train Predictive models Augmented with Isolated modality Ratings (PAIR) and demonstrate the potential of isolated modality ratings for enhancing multimodal emotion recognition. Our results suggest that perceptions of arousal are influenced primarily by auditory information, while perceptions of valence are more subjective and can be influenced by both visual and auditory information. The dataset is made publicly available.