Efforts to address declining accuracy as a result of data shifts often involve various data-augmentation strategies. Adversarial training is one such method, designed to improve robustness to worst-case distribution shifts caused by adversarial examples. While this method can improve robustness, it may also hinder generalization to clean examples and exacerbate performance imbalances across different classes. This paper explores the impact of adversarial training on both overall and class-specific performance, as well as its spill-over effects. We observe that enhanced labeling during training boosts adversarial robustness by 53.50 leading to improved accuracy in both clean and adversarial settings compared to standard adversarial training.

Non-robustness of image classifiers to subtle, adversarial perturbations is a well-known failure mode. Defenses against such attacks are typically evaluated by measuring the error rate on perturbed versions of the natural test set, quantifying the worst-case performance within a specified perturbation budget. However, these evaluations often isolate specific perturbation types, underestimating the adaptability of real-world adversaries who can modify or compose attacks in unforeseen ways. In this work, we show that models considered robust to strong attacks, such as AutoAttack, can be compromised by a simple modification of the weaker FGSM attack, where the adversarial perturbation is slightly transformed prior to being added to the input. Despite the attack's simplicity, robust models that perform well against standard FGSM become vulnerable to this variant. These findings suggest that current defenses may generalize poorly beyond their assumed threat models and can achieve inflated robustness scores under narrowly defined evaluation settings.

While dynamic ridesharing has been extensively studied, there remains a significant research gap in exploring role flexibility within the many-to-many ridesharing scheme, where the system allows for several pickups for drivers and multiple transfers for riders. Previous works have predominantly assumed that all participants own a car and have focused on one-to-one arrangements. Additionally, there is a scarcity of research on integrating High Occupancy Vehicle (HOV) lanes and mathematical modelling. This study addresses these gaps by presenting a novel Mixed Integer Linear Programming (MILP) model that allows for role flexibility irrespective of car ownership and considers the implications of HOV lanes. Computational analysis highlights the benefits of incorporating role flexibility and accommodating non-car-owning participants in many-to-many ridesharing systems. Yet, excessive role shifts may create imbalances, impacting service to non-car owners. Further research should explore these correlations.

Out-of-distribution generalization can be categorized into two types: common perturbations arising from natural variations in the real world and adversarial perturbations that are intentionally crafted to deceive neural networks. While deep neural networks excel in accuracy under the assumption of identical distributions between training and test data, they often encounter out-of-distribution scenarios resulting in a significant decline in accuracy. Data augmentation methods can effectively enhance robustness against common corruptions, but they typically fall short in improving robustness against adversarial perturbations. In this study, we develop Label Augmentation (LA), which enhances robustness against both common and intentional perturbations and improves uncertainty estimation. Our findings indicate a Clean error rate improvement of up to 23.29% when employing LA in comparisons to the baseline. Additionally, it enhances robustness under common corruptions benchmark by up to 24.23%. When tested against FGSM and PGD attacks, improvements in adversarial robustness are noticeable, with enhancements of up to 53.18% for FGSM and 24.46% for PGD attacks.

Medical imaging diagnosis increasingly relies on Machine Learning (ML) models. This is a task that is often hampered by severely imbalanced datasets, where positive cases can be quite rare. Their use is further compromised by their limited interpretability, which is becoming increasingly important. While post-hoc interpretability techniques such as SHAP and LIME have been used with some success on so-called black box models, the use of inherently understandable models makes such endeavours more fruitful. This paper addresses these issues by demonstrating how a relatively new synthetic data generation technique, STEM, can be used to produce data to train models produced by Grammatical Evolution (GE) that are inherently understandable. STEM is a recently introduced combination of the Synthetic Minority Oversampling Technique (SMOTE), Edited Nearest Neighbour (ENN), and Mixup; it has previously been successfully used to tackle both between-class and within-class imbalance issues. We test our technique on the Digital Database for Screening Mammography (DDSM) and the Wisconsin Breast Cancer (WBC) datasets and compare Area Under the Curve (AUC) results with an ensemble of the top three performing classifiers from a set of eight standard ML classifiers with varying degrees of interpretability. We demonstrate that the GE-derived models present the best AUC while still maintaining interpretable solutions.

The Connected Max-k-Cut Problem is an extension of the well-known Max-Cut Problem. The objective is to partition a graph into k connected subgraphs by maximizing the cost of inter-partition edges. We propose a new integer linear program for the problem and a branch-and-cut algorithm. We also explore graph isomorphism to structure the instances and facilitate their resolution. We conduct extensive computational experiments on both randomly generated instances and instances from the literature where we compare the quality of our method against existing algorithms. The experimental results show that, if k > 2, our approach strictly outperforms those from the literature. (c) 2023ElsevierB.V. Allrightsreserved.

Ensuring the secure and dependable deployment of deep neural networks hinges on their ability to withstand distributional shifts and distortions. While data augmentation enhances robustness, its effectiveness varies across different types of data corruption. It tends to excel in cases where corruptions share perceptually similar traits or have a high-frequency nature. In response, a strategy is to encompass a broad spectrum of distortions. Yet, it is often impractical to incorporate every conceivable modification that images may undergo within augmented data. Instead, we show that providing the model with a stronger inductive bias to learn the underlying concept of "change" would offer a more reliable approach. To this end, we develop Virtual Fusion (VF), a technique that treats corruptions as virtual labels. Diverging from conventional augmentation, when an image undergoes any form of transformation, its label becomes linked with the specific name attributed to the distortion. The finding indicates that VF effectively enhances both clean accuracy and robustness against common corruptions. On previously unseen corruptions, it shows an 11.90% performance improvement and a 12.78% increase in accuracy. In similar corruption scenarios, it achieves a 7.83% performance gain and a significant accuracy improvement of 22.04% on robustness benchmarks.

Memory phenotype of human GBM TIL

Co-expression of immune checkpoints in murine glioma

T cell exhaustion in brain metastases models

Cytokine production across immunologic compartments

Immune checkpoint and cytokine expression on human GBM CD4+ TIL

T cell counts per gram of tumor across tumor models

: Neuroevolution is the process of building or enhancing neural networks through the use of an evolutionary al-gorithm. An improved model can be defined as improving a model’s accuracy or finding a smaller model with faster training time with acceptable performance. Neural network hyper-parameter tuning is costly and time-consuming and often expert knowledge is required. In this study we investigate various methods to increase the performance of evolution, namely, epoch early stopping, using both improvement and threshold validation accuracy to stop training bad models, and removing duplicate models during the evolutionary process. Our results demonstrated the creation of a smaller model, 7 . 3 M , with higher accuracy, 0 . 969, in comparison to previously published methods. We also benefit from an average time saving of 59% because of epoch optimisation and 51% from the removal of duplicated individuals, compared to our prior work.

Gating strategy for tumor infiltrating lymphocytes

Deep learning is an effective and efficient method for image segmentation. Several neural network designs have been investigated, a notable example being the U-Net which has outperformed other segmentation models in different challenges. The Spatial Attention U-Net is a variation of the U-Net, which utilizes DropBlock and an attention block in addition to the typical U-Net convolutional blocks, which boosted the accuracy of the U-Net and reduced over-fitting. Optimising neural networks is costly, time-consuming and often requires expert guidance to determine the best mix of hyper-parameters for a particular problem. We demonstrate that grammatical evolution (GE) can be used to create U-Net and Spatial Attention U-Net architectures and optimise its choice of hyper-parameters. Our results show improved performance over state-of-the-art models on the Retinal Blood Vessel problem, increasing both AUC, from 0.978 to 0.979 and Accuracy, from 0.964 0.966, from the base models. Crucially, GE can achieve these improvements while finding a model which is 10 times smaller than the base models. A smaller model would enable its use in smart devices, such as smart phones, or in edge computing.

Patient and Control Characteristics

We study the problem of partitioning a graph into k connected components, which may also be referred to as the maximum k-cutset problem. Firstly, we present an exact algorithm and a variant, both implemented as integer linear programming (ILP) models. We then present a heuristic approach that will be seen to be extremely competitive with the exact algorithm for the ranges of graph under consideration.

Given a connected unit-weighted graph, we study the Maximum k-Cutset Problem, consisting in cutting a graph into k vertex-disjoint sub-graphs, each connected. We propose some exact and heuristic solutions to solve the problem. The first Integer Linear Program is based on a combination of a cut model and an assignment model. The two other Integer Linear Programs are based on the existence of spanning trees. The presented heuristics are clustering algorithms using local search. Experiments have been run on randomly generated instances and a specific set of instances. We observe that the running times are related to the tree-arboricity of graphs. The tree-arboricity is also used to ensure the optimality of solutions found by our heuristics.