SBIR Phase I: Geometric Unified Learning

这个小型企业创新研究(SBIR)第一阶段项目的更广泛的影响是，拟议的几何统一学习(GUL)技术，由可重复使用的构建块组成，克服了深度学习方法的几个缺点，并通过以下方式为最终用户带来了更大的价值：1)节省了数据处理的时间和资源，因为GUL具有自动数据清洗和压缩能力；2)在将数据输入到深度神经网络之前进行拟合，以避免过度拟合，因为GUL可以同时搜索和学习，并选择最重要的数据来使用；3)提供透明和可靠的解决方案，数据到决策者很容易理解；4)使机器学习领域以外的软件开发人员、科学家和工程师对数据分析感兴趣，因为GUL的输出是可理解的；5)由于GUL的组件可以被分解或重新组织以形成针对不同问题的新的解决方案，因此产生灵活、健壮和敏捷的解决方案用于调试；6)节省资金，因为GUL并行运行速度更快，仅使用约1%的原始数据点。这个小型企业创新研究(SBIR)第一阶段项目具有几个智力优势，可以解决深度学习中的几个技术挑战，例如产生垃圾输出，需要大量训练数据，运行时间长，无法解释的输出，过度的参数调整，然后重复，直到获得预期的结果。GUL植根于微分几何，创建适当的局部(GUL)坐标系、黎曼(GUL)度量、变换和测地线，以识别数据不变量、设计和设计GUL数据适配器的内在模式、特征抽取器和测量，每次最终用户运行他们的数据时，这些测量都可以得到加强，使其变得更好。GUL工具具有数据矢量化、数据压缩、搜索和学习同时进行的能力，并具有高度可解释性的结果。当这项技术做出预测时，它将向用户显示哪些数据点对这些预测负责。这一奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

The broader impact of this Small Business Innovation Research (SBIR) Phase I project is that the proposed Geometric Unified Learning (GUL) technology, consisting of reusable building blocks, overcomes several disadvantages of deep learning methods and brings increased value to end-users by: 1) saving time and resources in data processing, since GUL has an auto data cleaning and compressing capability; 2) fitting data before it is input to a deep neural network to avoid overfitting since GUL can search and learn simultaneously and select the most important data to use; 3) providing transparency and trustworthy solutions that can easily be understood by data-to-decision makers; 4) making data analysis interesting for software developers, scientists and engineers outside of the machine learning field since GUL outputs are comprehensible; 5) producing flexible, robust and agile solutions for debugging since GUL’s components can be decomposed or re-organized to form new solutions for different problems; and 6) saving money since GUL runs faster in parallel and only uses around 1% of the original data points.This Small Business Innovation Research (SBIR) Phase I project has several intellectual merits that address several technical challenges in deep learning, such as garbage-in producing garbage-out, needing large volumes of training data, long running times, unexplainable outputs, excessive parameter tuning, then repeating until the desired results are obtained. GUL, rooted in differential geometry, creates appropriate local (GUL) coordinate systems, Riemannian (GUL) metrics, transformations, and geodesics to identify data invariants, intrinsic patterns to engineer and design GUL data adaptors, characteristic extractors, and measurements that can be reinforced to be better each time an end-user runs their data. The GUL tools have capabilities of vectorizing data, compressing data, searching and learning simultaneously, with highly interpretable results. When the technology makes predictions it will show the user exactly which data points are responsible for those predictions.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.