Interpretable and Deblended Photometric Redshifts with a Deep Capsule Network

宇宙学、星系演化和星系团的研究都严重依赖于红移测量。然而，在当前和下一代深度成像调查中，只能获得一小部分星系的光谱红移，因此大多数红移必须仅从图像中估计。该项目引入了一种很有前途的新神经网络结构，称为深度胶囊网络，它将利用来自成像测量的像素级信息来估计光度学红移。主要目标是1)在通用的广域测试数据集上实现最先进的精度；2)利用遗留调查成像和暗能量光谱分析仪器数据将这些方法扩展到更高的红移；以及3)将分辨率光学成像与紫外线和红外的综合多波长光度测量相结合，重点关注可能存在引力波来源的极低红移星系。这项工作包括研究生和本科生的研究，以及夏季研究新兵训练营和每周的研讨会系列。训练营的所有课程材料都将公开提供，并开放访问。对于机器学习算法来说，估计光度红移是一个很好的问题，因为光谱红移可以为训练提供确定的测量结果。推动卷积神经网络广泛成功的汇集操作丢弃了细粒度的空间信息，并最终限制了它们概括为新观点和解析混合对象的能力。深胶囊网络克服了卷积神经网络的这些缺点，并且更容易解释。将开展的、适合学生研究的项目包括将斯隆胶囊网络结果与其他Photoz方法相结合以提高整体精度，以及将胶囊网络应用于模拟数据以专门测试混合对象的性能。该奖项反映了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。

Studies of cosmology, galaxy evolution, and galaxy clustering all critically depend on redshift measurements. However, spectroscopic redshifts can only be obtained for a small fraction of the galaxies detected in current and next-generation deep imaging surveys, so most redshifts must be estimated from the images alone. This project introduces a promising new neural network architecture, called a deep capsule network, which will leverage pixel-level information from imaging surveys to estimate photometric redshifts. Principal objectives are 1) achieving state-of-the-art accuracy on a common wide-field test data set; 2) extending these methods to higher redshifts using Legacy Survey imaging and Dark Energy Spectroscopic Instrument data; and 3) combining the resolved optical imaging with integrated multiwavelength photometry in the ultraviolet and infrared, focusing on very low redshift galaxies that might host gravitational wave sources. The work includes graduate and undergraduate research, and a summer research boot camp and weekly seminar series. All course materials from the bootcamp will be made publicly available with open access.Estimating photometric redshifts is a well-posed problem for machine learning algorithms because spectroscopic redshifts can provide definitive measurements for training. The pooling operation that fueled the widespread success of convolutional neural networks throws away fine-grained spatial information and ultimately limits their ability to generalize to novel viewpoints and to parse blended objects. The deep capsule network overcomes these drawbacks of convolutional neural networks, and is much more easily interpreted. Projects to be carried out, and suitable for student research, include combining Sloan capsule network results with other photo-z methods to improve overall accuracy, and applying capsule networks to simulated data specifically to test performance on blended objects.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.