Interpretable Learning Models Research Articles

Gaussian assumptions-free interpretable linear discriminant analysis for locating informative frequency bands for machine condition monitoring

Interpretable learning models become an emerging topic in the domain of machine condition monitoring to connect signal processing algorithms with statistical learning and machine learning. Compared with traditional signal processing algorithms that show strong ability in data analysis, signal-processing-related interpretable learning models could generate interpretable learnable weights/parameters as advanced physically interpretable fault features for both machine condition monitoring and fault diagnosis. It is well-known that linear discriminant analysis (LDA) is one of the most popular and interpretable algorithms for machine condition monitoring. However, this popular algorithm needs Gaussian assumptions in their derivations and parameter estimations. In this paper, Gaussian assumptions-free interpretable LDA is proposed as an interpretable learning model to physically locate informative frequency bands and fault characteristic frequencies for machine condition monitoring. Firstly, statistical decision theory is introduced to connect the nature of regression with that of classification, which poses a foundation for the Gaussian assumptions-free interpretable LDA for machine condition monitoring. Secondly, two propositions are given to mathematically show that LDA can be realized by an equivalent linear regression analysis, which provides a perspective for the Gaussian assumptions-free interpretable LDA for simultaneous machine condition monitoring and fault diagnosis. Finally, linear regression analysis with a sparse Lp-norm regularization term is introduced to realize the Gaussian assumptions-free interpretable LDA for physically locating informative frequency bands and fault characteristic frequencies for machine condition monitoring. Two case studies are provided as illustrative examples to experimentally demonstrate that the Gaussian assumptions-free interpretable LDA is capable of indicating informative frequency bands and fault characteristic frequencies.

Interpreting learning models in manufacturing processes: Towards explainable AI methods to improve trust in classifier predictions

Smart manufacturing processes, building upon machine learning (ML) models could potentially reduce the pre-production testing and validation time for new processes. Beyond calculating accurate and reliable models, one critical challenge would be for users of these models (plant operators, engineers and technicians) to trust these models’ outputs. We propose to apply explainable AI methods to create trustworthy AI-based manufacturing systems. Consequently, these systems will be enriched with capabilities to explain their reasoning processes and outputs (e.g., predictions) automatically. This paper applies explainable AI methods to two problems in manufacturing: ultrasonic weld (USW) quality prediction and body-in-white (BIW) dimensional variability reduction. Class activation maps were computed to explain the effect of input signals and their patterns on the quality predictions of an ultrasonic weld yield by a neural network (good or bad). Contrastive gradient based saliency maps were also created to assess the robustness of this classifier. Furthermore, we explain a given connectionist network that predicts the dimensional quality of body-in-white framer points based on deviations in underbody points. Explaining these predictions help engineers understand which underbody points have more influence on deviations in the framer points. These two applications highlight the importance of explainable AI methods in the modern manufacturing industry.

Fast Best Subset Selection: Coordinate Descent and Local Combinatorial Optimization Algorithms

In several scientific and industrial applications, it is desirable to build compact, interpretable learning models where the output depends on a small number of input features. Recent work has shown that such best-subset selection-type problems can be solved with modern mixed integer optimization solvers. Despite their promise, such solvers often come at a steep computational price when compared with open-source, efficient specialized solvers based on convex optimization and greedy heuristics. In “Fast Best-Subset Selection: Coordinate Descent and Local Combinatorial Optimization Algorithms,” Hussein Hazimeh and Rahul Mazumder push the frontiers of computation for best-subset-type problems. Their algorithms deliver near-optimal solutions for problems with up to a million features—in times comparable with the fast convex solvers. Their work suggests that principled optimization methods play a key role in devising tools central to interpretable machine learning, which can help in gaining a deeper understanding of their statistical properties.

AMRNN: attended multi-task recurrent neural networks for dynamic illness severity prediction

Illness severity prediction (ISP) is crucial for caregivers in the intensive care unit (ICU) while saving the life of patients. Existing ISP methods fail to provide sufficient evidence for the time-critical decision making in the dynamic changing environment. Moreover, the correlated temporal features in multivariate time-series are rarely be considered in existing machine learning-based ISP models. Therefore, in this paper, we propose a novel interpretable analysis framework which simultaneously analyses organ systems differentiated based on the pathological and physiological evidence to predict illness severity of patients in ICU. It not only timely but also intuitively reflects the critical conditions of patients for caregivers. In particular, we develop a deep interpretable learning model, namely AMRNN, which is based on the Multi-task RNNs and Attention Mechanism. Physiological features of each organ system in multivariate time series are learned by a single Long-Short Term Memory unit as a dedicated task. To utilize the functional and temporal relationships among organ systems, we use a shared LSTM task to exploit correlations between different learning tasks for further performance improvement. Real-world clinical datasets (MIMIC-III) are used for conducting extensive experiments, and our method is compared with the existing state-of-the-art methods. The experimental results demonstrated that our proposed approach outperforms those methods and suggests a promising way of evidence-based decision support.

Knowledge description for the suitability requirements of different geographical regions for growing wine

The production of wine has progressed on every main continent. The knowledge modeling can support the sharing of expertise, methods and good practice concerning international grape vine growing and wine production while maintaining a high level of quality. Our research focuses specifically on the development of a support system for knowledge formalization. We describe some procedural rules to represent experienced knowledge in the viticulture domain and plant pathology. We use a graphical software for rules management. The visual representation is a step toward the improvements of interaction between Artificial Intelligence methods and domain experts to make interpretable learning models for concrete decisions. This implementation enables us to make valuable visual reasoning to search whether the Chinese regions are capable of receiving a production of French vineyards. In particular, one outcome is that two Chinese regions appear more favorable and consistent for the development of wine from the Bordeaux region.