Transfer Learning for Time Domain Astrophysics

据预测，LSST每晚可以发现约1000万个瞬态和可变天体。这些物体的准确分类对于进行人口研究和识别稀有物体至关重要。我们建议从简单地使用光度数据转向利用光谱学可以提供的附加信息。光谱学允许对天体物理源的化学、物理和运动条件进行直接探测。然而，光谱学数据的收集比简单的光度法要耗费更多的时间。最近使用传统统计方法（如PCA）和更新颖的机器学习方法进行的实验表明，使用更先进的技术可以对低信噪比的观测结果进行分类。博士项目建议开发和使用更合适的机器学习方法，从使用深度学习（DL）进行降维开始。然而，可解释性问题是深度学习方法固有的性质，目前尚不清楚如何将先前的人类知识专长纳入深度学习模型。通过假设知识仅仅是从训练数据中提取出来的，我们忽略了宝贵的科学家经验。我们计划通过探索深度高斯过程（DGPs）的使用来解决这个问题。这些可以被描述为没有层内连接的深度学习网络，其中层之间的转换是概率的，并用高斯过程建模。dgp允许开发更多可解释的模型，并且可以将专家知识合并为先验分布。这项研究的成果对天体物理学和机器学习都有很大的潜在影响。在天体物理学中，低信噪比光谱的可分类性的改进允许在同一时间开发更大的样本量，并提高早期发现独特物体的机会，这通常是极端环境中物理理论的最佳测试。对于应用数学，它将推进目前对DGP的研究，DGP是一种仍处于发展阶段的方法，处于机器学习研究的前沿。

LSST is predicted to discover ~10 million transient and variable objects per night. Accurate classification of these objects will be crucial to allow both population studies and identify rare objects. We propose to move from simply using the photometric data to exploiting the additional information that can be afforded by spectroscopy. Spectroscopy allows a direct probe of the chemical, physical and kinematic conditions in an astrophysical source. However the collection of spectroscopic data is much more time consuming than simple photometry. Recent experiments using traditional statistical methods such as PCA and more novel machine learning approaches have shown some promise that low signal-to-noise ratio observations may be classifiable using more advanced techniques. The PhD project proposes the development and use of more suitable machine learning approaches starting with the use of deep learning (DL) for dimensionality reduction. However interpretability issues are inherent to the nature of DL methods and at present it is not clear how to incorporate prior human knowledge expertise into DL models. By assuming the knowledge is extracted solely from training data, we ignore valuable scientist experience. We plan to address this by exploring the use of Deep Gaussian Processes (DGPs). These can be characterized as deep learning networks without intra-layer connections, where the transformation between layers is probabilistic and modelled with Gaussian processes. DGPs could allow for developing more interpretable models, and the expert knowledge can be incorporated as prior distributions. The output of this research has a strong potential impact in both astrophysics and machine learning. In astrophysics, an improvement in classifiability of low signal to noise ratio spectra allows greater sample sizes to be developed in the same time and improves the chances of early detection of unique objects which can often be the best test of physical theories in extreme environments. For applied mathematics, it will advance the current research into DGP's, a method that is still in development and at the forefront of machine learning research.