Federated learning (FL), as an emerging artificial intelligence (AI) approach, enables decentralized model training across multiple devices without exposing their local training data. FL has been increasingly gaining popularity in both academia and industry. While research works have been proposed to improve the fault tolerance of FL, the real impact of unreliable devices (e.g., dropping out, misconfiguration, poor data quality) in real-world applications is not fully investigated. We carefully chose two representative, real-world classification problems with a limited numbers of clients to better analyze FL fault tolerance. Contrary to the intuition, simple FL algorithms can perform surprisingly well in the presence of unreliable clients.

Structural magnetic resonance imaging (MRI) studies demonstrated that Alzheimer's Disease (AD) causes not only local but also whole-brain level neural degenerative changes. To assess such changes, convolutional neural networks (CNN) are a popular approach as they are very capable automated feature extractors. In this work, due to the lack of segmentation that highlights brain degenerative changes, CNN-based brain age pre-diction is used as a surrogate task for training a feature extractor. Using the 3D T1-weighted MRI data of cognitive normal (CN) subjects from the OASIS-3 dataset, lightweight 3D CNN-based models are trained to predict brain age. The extracted features are then used in the binary classification of CN vs AD patients from their brain MRI scans. To established a baseline, we used support vector machines and random forest classifier as base classifiers. Our results suggest that the 3D MRI driven CNN brain age prediction surrogate task approach can learn AD-relevant features with high discriminative power without a complicated pipeline of preprocessing or data augmentation. Highlighting the novelty of the approach: the train-test-split is carefully performed on subject-level to avoid data leakage; the spatial information of 3D volumetric data is fully utilised; the robustness is proven by not using complicated data preprocessing and augmentation techniques.

Abstract This paper is a comprehensive survey of datasets for surgical tool detection and related surgical data science and machine learning techniques and algorithms. The survey offers a high level perspective of current research in this area, analyses the taxonomy of approaches adopted by researchers using surgical tool datasets, and addresses key areas of research, such as the datasets used, evaluation metrics applied and deep learning techniques utilised. Our presentation and taxonomy provides a framework that facilitates greater understanding of current work, and highlights the challenges and opportunities for further innovative and useful research.

We are introducing a new variation of the existing autoencoder called Radial Basis Function Autoencoders (RBFA). This version employs radial symmetric functions, in the first step of encoding, to map the input data vectors into a new form. The transformation which relies on the similarity between the input data examples and predefined center points features forces neurons in the first hidden level to respond similarly to similar patterns of the data. This basic idea of radial transformation (activation) can be interpreted as regularization on the hidden neurons which are penalized based on the similarity criterion so only a fraction of them will become highly activated at one time. The paper also introduces a new method for defining Gaussian radial function's parameters using Particle Swarm Optimization (PSO). Results indicate that the accuracy of four different classifiers was significantly affected by the number of layers in the first hidden level of the network. RBFA has shown promising results compared to other state of the art autoencoders using different datasets including high-dimensional small-sized medical data. Optimizing the radial function parameters improved the classification accuracy of the four classifiers in some cases.

Can evolutionary algorithms discover transformation functions capable of normalising arbitrary data? Many statistical and machine learning techniques that input data is normally distributed. However, in most realistic situations the data is distinctly non-normal. Therefore predictive performance may be compromised if steps are not taken to address this issue. We propose in this paper a new algorithm in which differential evolution is used to optimise the multiple parameters of a complex data transformation function. In order to optimise the transformation function, differential evolution is guided by a statistic used to measure the normality of the resulting output data. A set of soft constraints for controlling the scale and position of the output distribution are also included. On a set of fifteen datasets which includes a mixture of artificial data drawn from various different non-normal distributions; blood glucose level data from patients with diabetes; and gene expression data; our method is shown to effectively normalise the data in most cases.

Neural networks are powerful tools used widely for building cancer prediction models from microarray data. We review the most recently proposed models to highlight the roles of neural networks in predicting cancer from gene expression data. We identified articles published between 2013-2018 in scientific databases using keywords such as cancer classification, cancer analysis, cancer prediction, cancer clustering and microarray data. Analyzing the studies reveals that neural network methods have been either used for filtering (data engineering) the gene expressions in a prior step to prediction; predicting the existence of cancer, cancer type or the survivability risk; or for clustering unlabeled samples. This paper also discusses some practical issues that can be considered when building a neural network-based cancer prediction model. Results indicate that the functionality of the neural network determines its general architecture. However, the decision on the number of hidden layers, neurons, hypermeters and learning algorithm is made using trail-and-error techniques.

The goal of this research is to develop a lower-extremity wearable link mechanism (i.e., exoskeleton robot) that is capable of reducing load against targeted body parts such as bones, joints and muscles, for shock absorption that help to support exploration of extreme environments. One of the applications of such exoskeleton is to protect a pilot of a personal vertical mobility system, or JetPack, when landing. The shock absorbing exoskeleton is to introduce series and parallel viscoelasticity to the human skeletal system. The paper presents a pilot study to validate this body-protective exoskeleton concept by analyzing kinematic and dynamic models of a human-exoskeleton coupled system based on a multi-element viscoelastic model in rheology. A proof-of-concept prototype is developed and experimental data is presented.

A novel search operator, BlockCopy, is proposed for efficiently solving the wind farm layout optimisation problem. BlockCopy, which can be used either as mutation or a crossover operator, copies patterns of turbines from part of a layout to another part. The target layout may be the same as the source, or a different layout altogether. The rationale behind this is that it is the relative configurations of turbines rather than their individual absolute positions on the layouts that count, and BlockCopy, for the most part, maintains relative configurations. Our evaluation on four benchmark scenarios shows that BlockCopy outperforms two other standard approaches (namely, the turbine displacement algorithm and random perturbation) from the literature. We also evaluate the BlockCopy operator in conjunction with both single-solution and population-based strategies.

A novel mutation operator for the wind farm layout optimisation problem is proposed and tested. When a wind farm layout is simulated, statistics such as an individual turbine's wake free ratio can be computed. These statistics are in addition to the global measure being optimised, for example the overall cost of energy extraction of the farm. We present algorithms that first of all build a predictive model of the wake free ratio across an entire wind farm. This model is then used inside a mutation operator to perturb turbines towards positions of high predicted wake free ratio. We evaluate our approach by comparing a 1+1 Evolutionary Strategy using this new mutation operator vs. The same algorithm with a more standard random mutation operator, and show that our new operator leads to the discovery of wind farm layouts having a statistically significantly lower cost of energy extraction.

Differential Evolution can be used to construct effective and compact artificial training datasets for machine learning algorithms. In this paper, a series of comparative experiments are performed in which two simple interpretable supervised classifiers (specifically, Naive Bayes and linear Support Vector Machines) are trained (i) directly on “real” data, as would be the normal case, and (ii) indirectly, using special artificial datasets derived from real data via evolutionary optimization. The results across several challenging test problems show that supervised classifiers trained indirectly using our novel evolution-based approach produce models with superior predictive classification performance. Besides presenting the accuracy of the learned models, we also analyze the sensitivity of our artificial data optimization process to Differential Evolution's parameters, and then we examine the statistical characteristics of the artificial data that is evolved.

The introduction of next-generation DNA sequencing devices into the field of oncology provides an unprecedented mechanism to determine the underlying genetic changes that have occurred within a tumor and also the changes that accrue during treatment. An enhanced understanding of the oncogenic mechanisms could have an immediate clinical role in the treatment of rare tumors - where treatment protocols do not exist and their rarity would indicate that clinical trials would be unlikely to be undertaken for their establishment. We have investigated the utility of massively parallel sequencing to characterize a rare adenocarcinoma of the tongue, before and after treatment. In the pre-treatment tumor we identified 7,629 genes within regions of copy number gain, 1,078 genes exhibited increased expression relative to the blood and unrelated tumors and four genes contained somatic protein-coding mutations. Our analysis suggested the tumor cells were driven by the RET oncogene and its other pathway constituents. Genes whose protein products are targeted by the RET inhibitors sunitinib and sorafenib correlated with being amplified and or highly expressed. Consistent with our observations subsequent administration of sunitinib was associated with stable disease lasting 4 months, after which the lung lesions began to grow. Administration of sorafenib and sulindac provided disease stabilization for an additional 3 months after which the cancer progressed and new lesions appeared. A metastasis recurring in the skin was determined to possess 7,288 genes within copy number amplicons, 385 genes exhibiting increased expression relative to other tumours and 9 new somatic protein coding mutations. The observed mutations and amplifications were found to be consistent with resistance to therapy arising through further activation of RET pathway and nascent activation of the AKT pathway. Our results provide evidence for the clinical utility of complete genomic characterization and direct in-vivo genome-wide characterization of the mutations accruing within a tumor under drug selection.

A novel algorithm for 3D face recognition based point cloud rotations, multiple projections, and voted keypoint matching is proposed and evaluated. The basic idea is to rotate each 3D point cloud representing an individual's face around the x, y or z axes, iteratively projecting the 3D points onto multiple 2.5D images at each step of the rotation. Labeled keypoints are then extracted from the resulting collection of 2.5D images, and this much smaller set of keypoints replaces the original face scan and its projections in the face database. Unknown test faces are recognized firstly by performing the same multiview keypoint extraction technique, and secondly, the application of a new weighted keypoint matching algorithm. In an extensive evaluation using the GavabDB 3D face recognition dataset (61 subjects, 9 scans per subject), our method achieves up to 95% recognition accuracy for faces with neutral expressions only, and over 90% accuracy for face recognition where expressions (such as a smile or a strong laugh) and random face-occluding gestures are permitted.