On the Robustness of Bayesian Neural Networks to Adversarial Attacks.

Vulnerability to adversarial attacks is one of the principal hurdles to the adoption of deep learning in safety-critical applications. Despite significant efforts, both practical and theoretical, training deep learning models robust to adversarial attacks is still an open problem. In this article, we analyse the geometry of adversarial attacks in the over-parameterized limit for Bayesian neural networks (BNNs). We show that, in the limit, vulnerability to gradient-based attacks arises as a result of degeneracy in the data distribution, i.e., when the data lie on a lower dimensional submanifold of the ambient space. As a direct consequence, we demonstrate that in this limit, BNN posteriors are robust to gradient-based adversarial attacks. Crucially, by relying on the convergence of infinitely-wide BNNs to Gaussian processes (GPs), we prove that, under certain relatively mild assumptions, the expected gradient of the loss with respect to the BNN posterior distribution is vanishing, even when each NN sampled from the BNN posterior does not have vanishing gradients. The experimental results on the MNIST, Fashion MNIST, and a synthetic dataset with BNNs trained with Hamiltonian Monte Carlo and variational inference support this line of arguments, empirically showing that BNNs can display both high accuracy on clean data and robustness to both gradient-based and gradient-free adversarial attacks.

L-lactate production in engineered Saccharomyces cerevisiae using a multistage multiobjective automated design framework.

The design of alternative biodegradable polymers has the potential of severely reducing the environmental impact, cost and production time currently associated with the petrochemical industry. In fact, growing demand for renewable feedstock has recently brought to the fore synthetic biology and metabolic engineering. These two interdependent research areas focus on the study of microbial conversion of organic acids, with the aim of replacing their petrochemical-derived equivalents with more sustainable and efficient processes. The particular case of Lactic acid (LA) production has been the subject of extensive research because of its role as an essential component for developing an eco-friendly biodegradable plastic-widely used in industrial biotechnological applications. Because of its resistance to acidic environments, among the many LA-producing microbes, Saccharomyces cerevisiae has been the main focus of research into related biocatalysts. In this study, we present an extensive in silico investigation of S. cerevisiae cell metabolism (modeled with Flux Balance Analysis) with the overall aim of maximizing its LA production yield. We focus on the yeast 8.3 steady-state metabolic model and analyze it under the impact of different engineering strategies including: gene knock-in, gene knock-out, gene regulation and medium optimization; as well as a comparison between results in aerobic and anaerobic conditions. We designed ad-hoc constrained multiobjective evolutionary algorithms to automate the engineering process and developed a specific postprocessing methodology to analyze the genetic manipulation results obtained. The in silico results reported in this paper empirically show that our method is able to automatically select a small number of promising genetic and metabolic manipulations, deriving competitive strains that promise to impact microorganisms design in the production of sustainable chemicals.

Physiologically-Informed Gaussian Processes for Interpretable Modelling of Psycho-Physiological States.

The widespread popularity of Machine Learning (ML) models in healthcare solutions has increased the demand for their interpretability and accountability. In this paper, we propose the Physiologically-Informed Gaussian Process (PhGP) classification model, an interpretable machine learning model founded on the Bayesian nature of Gaussian Processes (GPs). Specifically, we inject problem-specific domain knowledge of inherent physiological mechanisms underlying the psycho-physiological states as a prior distribution over the GP latent space. Thus, to estimate the hyper-parameters in PhGP, we rely on the information from raw physiological signals as well as the designed prior function encoding the physiologically-inspired modelling assumptions. Alongside this new model, we present novel interpretability metrics that highlight the most informative input regions that contribute to the GP prediction. We evaluate the ability of PhGP to provide an accurate and interpretable classification on three different datasets, including electrodermal activity (EDA) signals collected during emotional, painful, and stressful tasks. Our results demonstrate that, for all three tasks, recognition performance is improved by using the PhGP model compared to competitive methods. Moreover, PhGP is able to provide physiological sound interpretations over its predictions.

Current State and Future Directions of Technology-Based Ecological Momentary Assessment and Intervention for Major Depressive Disorder: A Systematic Review.

Ecological momentary assessment (EMA) and ecological momentary intervention (EMI) are alternative approaches to retrospective self-reports and face-to-face treatments, and they make it possible to repeatedly assess patients in naturalistic settings and extend psychological support into real life. The increase in smartphone applications and the availability of low-cost wearable biosensors have further improved the potential of EMA and EMI, which, however, have not yet been applied in clinical practice. Here, we conducted a systematic review, using the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines, to explore the state of the art of technology-based EMA and EMI for major depressive disorder (MDD). A total of 33 articles were included (EMA = 26; EMI = 7). First, we provide a detailed analysis of the included studies from technical (sampling methods, duration, prompts), clinical (fields of application, adherence rates, dropouts, intervention effectiveness), and technological (adopted devices) perspectives. Then, we identify the advantages of using information and communications technologies (ICTs) to extend the potential of these approaches to the understanding, assessment, and intervention in depression. Furthermore, we point out the relevant issues that still need to be addressed within this field, and we discuss how EMA and EMI could benefit from the use of sensors and biosensors, along with recent advances in machine learning for affective modelling.

Current state and future directions of technology-based ecological momentary assessments and interventions for major depressive disorder: protocol for a systematic review.

BACKGROUND: Ecological momentary assessments (EMAs) and ecological momentary interventions (EMIs) represent a novel approach for the assessment and delivery of psychological support to depressed patients in daily life. Beyond the classical paper-and-pencil daily diaries, the more recent progresses in Information and Communication Technologies (ICT) enabled researchers to bring all the needed processes together in only one device, i.e., response signaling, repeated symptom collection, information storage, secure data transfer, and psychological support delivery. Despite evidence showing the feasibility and acceptability of these techniques, EMAs are only beginning to be applied in real clinical practice, whether the development of EMIs for clinically depressed patients is still very limited. The objective of this systematic review is to provide the state of the art of technology-based EMAs and EMIs for major depressive disorder (MDD), with the aim of leading the way to possible future directions for the clinical practice. METHODS: We will conduct a systematic review using the Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA) guidelines. Data sources will include two bibliographic databases, PubMed and Web of Science (Web of Knowledge), supplemented by searches for unpublished or ongoing studies. Eligible studies will report data for adult (≥ 18 years old) with a primary (both current and past) diagnosis of MDD, defined by a valid criterion standard. We will consider studies adopting technology-based EMAs and EMIs for the investigation and/or assessment of depression and for the delivery of a psychological intervention. We will exclude studies adopting paper-and-pencil tools. DISCUSSION: The proposed systematic review will provide new insights on the advantages and benefits of adopting technology-based EMAs and EMIs for MDD in the traditional clinical practice, taking into consideration both clinical and technological issues. The potential of using sensors and biosensors along with machine learning for affective modeling will also be discussed.

Pareto Optimal Design for Synthetic Biology.

Recent advances in synthetic biology call for robust, flexible and efficient in silico optimization methodologies. We present a Pareto design approach for the bi-level optimization problem associated to the overproduction of specific metabolites in Escherichia coli. Our method efficiently explores the high dimensional genetic manipulation space, finding a number of trade-offs between synthetic and biological objectives, hence furnishing a deeper biological insight to the addressed problem and important results for industrial purposes. We demonstrate the computational capabilities of our Pareto-oriented approach comparing it with state-of-the-art heuristics in the overproduction problems of i) 1,4-butanediol, ii) myristoyl-CoA, i ii) malonyl-CoA , iv) acetate and v) succinate. We show that our algorithms are able to gracefully adapt and scale to more complex models and more biologically-relevant simulations of the genetic manipulations allowed. The Results obtained for 1,4-butanediol overproduction significantly outperform results previously obtained, in terms of 1,4-butanediol to biomass formation ratio and knock-out costs. In particular overproduction percentage is of +662.7%, from 1.425 mmolh⁻¹gDW⁻¹ (wild type) to 10.869 mmolh⁻¹gDW⁻¹, with a knockout cost of 6. Whereas, Pareto-optimal designs we have found in fatty acid optimizations strictly dominate the ones obtained by the other methodologies, e.g., biomass and myristoyl-CoA exportation improvement of +21.43% (0.17 h⁻¹) and +5.19% (1.62 mmolh⁻¹gDW⁻¹), respectively. Furthermore CPU time required by our heuristic approach is more than halved. Finally we implement pathway oriented sensitivity analysis, epsilon-dominance analysis and robustness analysis to enhance our biological understanding of the problem and to improve the optimization algorithm capabilities.

Hardware-in-the-loop simulation and energy optimization of cardiac pacemakers.

Implantable cardiac pacemakers are medical devices that can monitor and correct abnormal heart rhythms. To provide the necessary safety assurance for pacemaker software, both testing and verification of the code, as well as testing the entire pacemaker hardware in the loop, is necessary. In this paper, we present a hardware testbed that enables detailed hardware-in-the-loop simulation and energy optimisation of pacemaker algorithms with respect to a heart model. Both the heart and the pacemaker models are encoded in Simulink/Stateflow™ and translated into executable code, with the pacemaker executed directly on the microcontroller. We evaluate the usefulness of the testbed by developing a parameter synthesis algorithm which optimises the timing parameters based on power measurements acquired in real-time. The experiments performed on real measurements successfully demonstrate that the testbed is capable of energy minimisation in real-time and obtains safe pacemaker timing parameters.