SI2-SSE: Deep Forge: a Machine Learning Gateway for Scientific Workflow Design

机器学习的最新进展已经对我们的生活产生了变革性的影响。然而，在图像分类、语音识别、自动驾驶汽车和自然语言处理等不同领域取得的惊人成功，大多是由商业力量推动的，这些技术尚未广泛过渡到各种科学领域。该领域的创新时机已经成熟，因为许多科学领域都有现成的大规模数据集，以及能够执行计算成本高昂的人工神经网络（ANN）训练工作流程的公共或私人计算基础设施。主要的障碍似乎是人工神经网络工具的陡峭学习曲线，建立和执行机器学习工作流程的意外复杂性，以及找到正确的深度神经网络架构需要大量的经验和大量的实验。DeepForge通过提供直观的可视化界面、大型可重用组件和架构库以及自动软件生成来克服这些障碍，使领域科学家能够在自己的领域中实验人工神经网络。机器学习在各种各样的应用领域都有很大的潜力，因此对人才的需求很高。因此，任何帮助科学家在自己的领域应用机器学习的工具都将产生广泛的影响。DeepForge的承诺是使学习曲线变得平坦，隐藏低级不重要的细节，并提供可在学科内部和跨学科中重用的组件。因此，DeepForge将对许多领域产生变革性的影响。DeepForge是一个基于网络和云的软件基础设施，通过直观的视觉界面和管理训练工件，提高了创建人工神经网络工作流的抽象性。因此，它使领域科学家能够利用机器学习的最新进展。DeepForge还将集成现有的网络基础设施，包括私有和商业计算集群、云服务（例如亚马逊EC2）、公共超级计算资源和科学数据集的在线存储库。用于设计人工神经网络架构和工作流的DeepForge可视化语言足够强大，可以捕获与常见深度学习任务相关的概念，但它提供了一个高层次的抽象，同时保护用户免受潜在复杂性的影响。DeepForge将提供一个设施，允许在广泛的跨学科用户社区中共享设计工件。策划一个丰富的可重用组件库，与各种现有的网络基础设施资源集成，从数据源到计算平台，并以无缝的方式提供数据来源是该项目的其他优势。DeepForge将通过与数字对象标识符（DOI）基础设施的集成，推广“数据即产品”、“模型即产品”和“服务即产品”的概念。DeepForge将使科学家能够为他们的共享资产分配doi，提供数据来源，通过使用关联的数据工件执行引用的人工神经网络工作流，从而引用和公开复制研究结果。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Recent advances in machine learning have already had a transformative impact on our lives. However, astonishing successes in diverse domains, such as image classification, speech recognition, self-driving cars and natural language processing, have mostly been driven by commercial forces, and these techniques have not yet been widely transitioned into various science domains. The field is ripe for innovation since many science fields have readily available large-scale datasets, as well as access to public or private compute infrastructure capable of executing computationally expensive artificial neural network (ANN) training workflows. The main roadblocks seem to be the steep learning curve of the ANN tools, the accidental complexities of setting up and executing machine learning workflows, and the fact that finding the right deep neural network architecture requires significant experience and lots of experimentation. DeepForge overcomes these obstacles by providing an intuitive visual interface, a large library of reusable components and architectures as well as automatic software generation enabling domain scientist to experiment with ANNs in their own field. There is unmet high demand of talent in machine learning, exactly because it has so much potential in a wide variety of application areas. Therefore, any tool that helps scientists apply machine learning in their own domains will have a broad impact. The promise of DeepForge is to flatten the learning curve, hide low level unimportant details and provide components that are reusable within and across disciplines. Therefore, DeepForge will have transformative impact on a number of fields.DeepForge, a web- and cloud-based software infrastructure raises the abstraction of creating ANN workflows via an intuitive visual interface and by managing training artifacts. Hence, it enables domain scientists to leverage recent advances in machine learning. DeepForge will also integrate with existing cyberinfrastructure, including private and commercial compute clusters, cloud services (e.g. Amazon EC2), public supercomputing resources, and online repositories of scientific datasets. The DeepForge visual language for designing ANN architectures and workflows is powerful enough to capture the concepts related to common deep learning tasks, yet it provides a high level of abstraction that shields the users from the underlying complexity at the same time. DeepForge will provide a facility that allows for sharing design artifacts across a wide interdisciplinary user community. Curating a rich library of reusable components, integrating with a wide variety of existing cyberinfrastructure resources from data sources to compute platform and providing data provenance in a seamless manner are other advantages of the project. DeepForge will promote "data as product," "model as product," and "service as product" concepts through integration with the Digital Object Identifier (DOI) infrastructure. DeepForge will enable scientist to assign DOIs to their shared assets providing data provenance enabling citing and publicly reproducing research results by executing the referenced ANN workflows with the linked data artifacts.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.