David L. J. Alexander

Transparency in AI models is essential for promoting human–AI collaboration and ensuring regulatory compliance. However, interpreting these models is a complex process influenced by various methods and datasets. This study presents a comprehensive overview of foundational interpretation techniques, meticulously referencing the original authors and emphasizing their pivotal contributions. Recognizing the seminal work of these pioneers is imperative for contextualizing the evolutionary trajectory of interpretation in the field of AI. Furthermore, this research offers a retrospective analysis of interpretation techniques, critically evaluating their inherent strengths and limitations. We categorize these techniques into model-based, representation-based, post hoc, and hybrid methods, delving into their diverse applications. Furthermore, we analyze publication trends over time to see how the adoption of advanced computational methods within various categories of interpretation techniques has shaped the development of AI interpretability over time. This analysis highlights a notable preference shift towards data-driven approaches in the field. Moreover, we consider crucial factors such as the suitability of these techniques for generating local or global insights and their compatibility with different data types, including images, text, and tabular data. This structured categorization serves as a guide for practitioners navigating the landscape of interpretation techniques in AI. In summary, this review not only synthesizes various interpretation techniques but also acknowledges the contributions of their original authors. By emphasizing the origins of these techniques, we aim to enhance AI model explainability and underscore the importance of recognizing biases, uncertainties, and limitations inherent in the methods and datasets. This approach promotes the ethical and practical use of interpretation insights, empowering AI practitioners, researchers, and professionals to make informed decisions when selecting techniques for responsible AI implementation in real-world scenarios.

Machine-learning assisted optimisation of a continuous photodegradation reaction, using a TiO₂ coated catalytic static mixer successfully accounting for catalyst degradation.

Time-of-flight secondary ion mass spectrometry (ToF-SIMS) imaging is used across many fields for the atomic and molecular characterization of surfaces, with both high sensitivity and high spatial resolution. When large analysis areas are required, standard ToF-SIMS instruments allow for the acquisition of adjoining tiles, which are acquired by rastering the primary ion beam. For such large area scans, tiling artifacts are a ubiquitous challenge, manifesting as intensity gradients across each tile and/or sudden changes in intensity between tiles. Such artifacts are thought to be related to a combination of sample charging, local detector sensitivity issues, and misalignment of the primary ion gun, among other instrumental factors. In this work, we investigated six different computational tiling artifact removal methods: tensor decomposition, multiplicative linear correction, linear discriminant analysis, seamless stitching, simple averaging, and simple interpolating. To ensure robustness in the study, we applied these methods to three hyperspectral ToF-SIMS data sets and one OrbiTrapSIMS data set. Our study includes a carefully designed statistical analysis and a quantitative survey that subjectively assessed the quality of the various methods employed. Our results demonstrate that while certain methods are useful and preferred more often, no one particular approach can be considered universally acceptable and that the effectiveness of the artifact removal method is strongly dependent on the particulars of the data set analyzed. As examples, the multiplicative linear correction and seamless stitching methods tended to score more highly on the subjective survey; however, for some data sets, this led to the introduction of new artifacts. In contrast, simple averaging and interpolation methods scored subjectively poorly on the biological data set, but more highly on the microarray data sets. We discuss and explore these findings in depth and present general recommendations given our findings to conclude the work.

The self-organizing map (SOM) is a nonlinear machine learning algorithm that is particularly well suited for visualizing and analyzing high-dimensional, hyperspectral time-of-flight secondary ion mass spectrometry (ToF-SIMS) imaging data. Previously, we compared the capabilities of the SOM with more traditional linear techniques using ToF-SIMS imaging data. Although SOMs perform well with minimal data preprocessing and negligible hyperparameter optimization, it is important to understand how different data preprocessing methods and hyperparameter settings influence the performance of SOMs. While these investigations have been reported outside of the ToF-SIMS field, no such study has been reported for hyperspectral MSI data. To address this, we used two labeled ToF-SIMS imaging datasets, one of which was a polymer microarray dataset, while the other was semisynthetic hyperspectral data. The latter was generated using a novel algorithm that we describe here. A grid-search was used to evaluate which data preprocessing methods and SOM hyperparameters had the largest impact on the performance of the SOM. This was assessed using multiple linear regression, whereby performance metrics were regressed onto each variable defining the preprocessing-hyperparameter space. We found that preprocessing was generally more important than hyperparameter selection. We also found statistically significant interactions between several parameters studied, suggesting a complex interplay between preprocessing and hyperparameter selection. Importantly, we identified interesting trends, both dataset specific and dataset agnostic, which we describe and discuss in detail.

Many real-world optimisation problems are defined over both categorical and continuous variables, yet efficient optimisation methods such as Bayesian Optimisation (BO) are ill-equipped to handle such mixed-variable search spaces. The optimisation breadth introduced by categorical variables in the mixed-input setting has seen recent approaches operating on local trust regions, but these methods can be greedy in suboptimal regions of the search space. In this paper, we adopt a holistic view and aim to consolidate optimisation of the categorical and continuous sub-spaces under a single acquisition metric. We develop a tree-based method which retains a global view of the optimisation spaces by identifying regions in the search space with high potential candidates which we call value proposals. Our method uses these proposals to make selections on both the categorical and continuous components of the input. We show that this approach significantly outperforms existing mixed-variable optimisation approaches across several mixed-variable black-box optimisation tasks.

Certified disposable respirators afford important protection from hazardous aerosols but lose performance as they are worn. This study examines the effect of wear time on filtration efficiency. Disposable respirators were worn by CSIRO staff over a period of 4 weeks in early 2020. Participants wore the respirator masks for given times up to eight hours whilst working in laboratory/office environments. At that time COVID-19 precautions required staff to wear surgical (or other) masks and increase use of hand sanitizer from dispenser stations. Results obtained from a test group of ten individuals without health preconditions show an increasing number of masks failing with wear time, while the remainder continue to perform nearly unaffected for up to 8 h. Some masks were found to retain filtration performance better than others, possibly due to the type of challenge they were subjected to by the wearer. However, the rate and extent of decay are expected to differ between environments since there are many contributing factors and properties of the aerosol challenge cannot be controlled in a live trial. Penetration and variability increased during wear; the longer the wear time, the more deleterious to particle removal, particularly after approximately 2 h of wear. This behavior is captured in a descriptive statistical model based on results from a trial with this test group. The effectiveness of the masks in preventing the penetration of KCl particles was determined before and after wearing, with the analysis focusing on the most penetrating particles in a size range of 0.3–0.5 µm diameter where respirator masks are most vulnerable. The basic elements of the study, including the approach to filter testing and sample sanitization, are broadly applicable. Conclusions also have applicability to typical commercially available single-use respirator masks manufactured from melt blown polypropylene as they are reliant on the same physical principles for particle capture and electrostatic enhancement was comparable for the particle size range used for detection.

The potential for alcoholic vapors emitted by common sanitizing treatments to deteriorate the (electrostatic) filtration performance of disposable respirator masks has been investigated. Reports in the literature and some standard test methods provide a confusing and ambiguous picture concerning the relevance of this effect. Four different types of exposure were investigated in this study to clarify the effect of alcoholic vapor emissions on respirator masks. These included exposure to saturated vapors, use of hand sanitizers, cleaning of table surfaces and sanitization of masks by spraying them with alcohol-containing solutions. Methods employed were designed to be as real-world oriented as possible while remaining reproducible. Filtration performance and deterioration effects on exposure to the different treatments were determined on three different types of certified commercial respirator masks—a P2 and two KN95 masks. This study provides substantial evidence that disposable respirator masks with an accepted performance rating are seriously compromised from an exposure to saturated alcoholic vapors, can tolerate a one-off spray treatment with an alcoholic solution and retain their attested protection under the influence of alcoholic vapors from the use of hand sanitizer or spray sanitizer. Considering the range of vastly different outcomes obtained from the four treatments investigated, it seems prudent to assess in each case the specific effects of alcoholic solution treatments and vapors on respirator masks before use.