MiFL: Multi-Input Neural Networks in Federated Learning

Abstract

<p>Driven by the Deep Learning (DL) revolution, Artificial intelligence (AI) has become a fundamental tool for many Bio-Medical tasks, including AI-assisted diagnosis. These include analysing and classifying images (2D and 3D), where, for some tasks, DL exhibits superhuman performance. Diagnostic imaging, however, is not the only diagnostic tool. Tabular data, such as personal data, vital signs, and genomic/blood tests, are commonly collected for every patient entering a clinical institution. However, it is rarely considered in DL pipelines, although it carries diagnostic information. The training of DL models requires large datasets, so large that every institution might need more data that should be pooled from different sites. Data pooling generates newfound concerns about data access and movement across other institutions spawning multiple dimensions, such as performance, energy efficiency, privacy, criticality, and security. Federated Learning (FL) is a cooperative learning paradigm aiming at addressing these concerns by moving models instead of data across different institutions. This paper proposes a Federated multi-input model that leverages images and tabular data, providing a proof of concept of the feasibility of multi-input FL architectures. The proposed model was evaluated on two showcases: the prognosis of CoViD-19 disease and the patients’ stratification for Alzheimer’s disease. Results show that enabling multi-input architectures in the FL framework allows for improving the performance regarding both accuracy and generalizability with respect to non-federated models while ensuring security and data protection peculiar to FL.</p>

Experimenting With Normalization Layers in Federated Learning on Non-IID Scenarios

“…Results in [11] also confirmed literature sentiment: labels quantity skew and its pathological variant are the most detrimental ones for the algorithms. The same non-IID partitions have already been tested in [24], which proposes a novel technique of non-gradient-descent FL on tabular datasets.…”

“…The same non-IID partitions have already been tested in [24], which proposes a novel technique of non-gradient-descent FL on tabular datasets. Our paper extends [11], deepening the experiments about the number of epochs per round, a hyper-parameter that, if tuned appropriately, can lead to large performance gains. Moreover, we aim to investigate which type of normalization layer better fits FL on non-IID data.…”

“…Li et al [10] show that none of those algorithms outperforms others in all the cases and that non-iidness degrades the performance of FL systems in terms of accuracy, especially in the case of label quantity skew. Another recent work [11] reports an empirical assessment of the behavior of FedAvg and FedCurv [9] on MNIST, CIFAR10 and MedMNIST [20]. Datasets are split according to the same non-IID settings of [10].…”

“…The most common non-IID data settings are reviewed in [1] that lists five different partitioning strategies: 1) quantity skew, 2) labels quantity skew (prior shift), 3) feature distribution skew (covariate shift), 4) same labels but different features and 5) same features but different labels. In this paper, we consider the same distributions tested in [10], [11] apart from quantity skew, which is not treated. Indeed, [10], [11] showed that quantity skew does not hurt the performance of FL models, probably because it results in a different quantity of samples per client, but the distribution of samples is uniform, which is easy to deal with.…”

“…Recent works have proposed several FL algorithms to cope with non-IIDness problems, such as FedProx [6], Fed-Nova [7], SCAFFOLD [8], and FedCurv [9], which has been tested in [10], [11]. Notice that all these algorithms are modified versions of FedAvg, and they preserve the principle underneath FedAvg: to average the weights in all the layers of the NN.…”