Distributed Deep Learning using Blockchain Mining Servers for Medical Imaging

近年来，各种新的深度神经网络(DNN)体系结构被提出，如GoogLeNet和VGGNet**，以解决分类和目标识别等人工智能问题。更大的公共数据集的出现，如ImageNet，进一步刺激了人工智能的研究，并导致了**更深和更具表现力的DNN的设计，以便模型的复杂性足以完成任务。尽管**图形处理单元(GPU)为深度学习提供了巨大的计算能力，但这些**更深入、更复杂的模型需要在单个GPU上进行数周甚至数月的训练，这导致了**分布式GPU培训的研究。**根据这一需求，星云AI Inc.的子公司Service ECVictor Inc.部署了一种新型的分布式**计算架构，将区块链挖掘机(Miner)的散列能力转换为**GPU计算节点组(每个Miner包含多个GPU)。星云AI旨在提供一个高效、低**成本、安全可靠的计算平台来处理各种人工智能(AI)任务。对于**当前的Engage项目，星云AI将扩展其平台上可用的解决方案(主要基于**数字或基于文本的数据)，以包括基于深度学习模型的分布式训练的解决方案**使用2D和3D图像数据的方法，这在医疗保健的深度学习应用程序中很常见。为了**测试这些解决方案，这些解决方案将用于开发基于**糖尿病视网膜病变检查数据库的**深度学习的计算机辅助医疗诊断系统。由此产生的研究出版物将**展示星云AI平台用于医学成像深度学习**问题的可行性。

In recent years, a variety of new deep neural networks (DNN) architectures such as GoogleNet and VGGNet**have been proposed to solve AI problems like classification and object identification. The emergence of larger**public datasets, such as ImageNet, has further spurred artificial intelligence research and lead to the design of**deeper and more expressive DNNs so that the complexity of models is sufficient for the task. Although**graphics processing unites (GPU) provides enormous computational power for deep learning purposes, these**deeper and more complex models require weeks or even months to be trained on a single GPU which has led to**research in distributed GPU training.**In line with this need, Service ECVictor Inc, a subsidiary of Nebula AI Inc., has deployed a novel distributed**computing architecture that converts the hash power from blockchain mining machines (miners) into groups of**GPU computing nodes (each miner containing multiple GPUs). Nebula AI aims to offer a highly efficient, low**cost, safe and reliable computing platform to handle a wide variety of Artificial Intelligence (AI) tasks. For the**current Engage project, Nebula AI will expand the solutions available on their platform (primarily based on**numerical or text-based data) to include solutions based on the distributed training of models for deep learning**approaches using 2D and 3D image data, which are common in deep learning applications for healthcare. To**test these solutions, the solutions will be used to develop computer-aided medical diagnostic systems based on**deep learning for databases of Diabetic Retinopathy examinations. The resulting research publications will**demonstrate the viability of the Nebula AI platform to be used for Deep Learning for Medical Imaging**problems.