Machines Learn to Infer Stellar Parameters Just by Looking at a Large Number of Spectra

TL;DR

The paper tackles inferring stellar parameters from large, unlabeled spectral data by training a self-supervised convolutional autoencoder on HARPS spectra and analyzing its latent space. By enforcing disentanglement through a β-VAE–style objective and using mutual information and dispersion-based metrics, the authors identify latent dimensions that align with physical quantities such as radial velocity and effective temperature, and uncover additional informative features not directly tied to labeled labels. Key findings include the emergence of about six informative latent dimensions, with two nodes clearly representing RV and Teff, and evidence of latent-space structure via traversal experiments; results are robust to some data-balancing variations but sensitive to sampling biases. The approach demonstrates a data-driven pathway to uncover physical relationships in astronomy, offers a framework for discovering new patterns in large spectra datasets, and provides public code and an interactive interface to facilitate further science driven by learned representations.

Abstract

Machine learning has been widely applied to clearly defined problems of astronomy and astrophysics. However, deep learning and its conceptual differences to classical machine learning have been largely overlooked in these fields. The broad hypothesis behind our work is that letting the abundant real astrophysical data speak for itself, with minimal supervision and no labels, can reveal interesting patterns which may facilitate discovery of novel physical relationships. Here as the first step, we seek to interpret the representations a deep convolutional neural network chooses to learn, and find correlations in them with current physical understanding. We train an encoder-decoder architecture on the self-supervised auxiliary task of reconstruction to allow it to learn general representations without bias towards any specific task. By exerting weak disentanglement at the information bottleneck of the network, we implicitly enforce interpretability in the learned features. We develop two independent statistical and information-theoretical methods for finding the number of learned informative features, as well as measuring their true correlation with astrophysical validation labels. As a case study, we apply this method to a dataset of ~270000 stellar spectra, each of which comprising ~300000 dimensions. We find that the network clearly assigns specific nodes to estimate (notions of) parameters such as radial velocity and effective temperature without being asked to do so, all in a completely physics-agnostic process. This supports the first part of our hypothesis. Moreover, we find with high confidence that there are ~4 more independently informative dimensions that do not show a direct correlation with our validation parameters, presenting potential room for future studies.

Figures (14)

• Figure 1: A large number of stellar spectra are passed through the information bottleneck of a deep convolutional autoencoder, in a fully unsupervised, physics-agnostic process. The network has zero information about the content of the numerical vectors it receives. We use techniques based on information maximization, to enforce learning of disentangled features, and find that the network learns representations for astrophysical parameters such as radial velocity and effective temperature, without being asked to do so.
• Figure 2: Brief architecture of the deterministic autoencoder on top, with the schematic variational counterpart of it at the bottom. In the VAE version, the code is not directly connected to the encoder, but is drawn from the learnable parameters of the normal distribution: reparametrization trickkingma_auto-encoding_2014.
• Figure 3: Illustration of the effects of two major factors on reconstruction quality: latent space dimensionality and disentanglement. The left two columns illustrate reconstruction loss over the whole spectra, while on the right the same effects are depicted, in two different zoom levels, on an exemplar single spectrum: Input(blue) and reconstructed version(orange) are overplotted. Comparing the results of the deterministic autoencoder, and that of the disentangled variational autoencoder, we can clearly see the sacrifice in reconstruction quality, that occurs for the sake of disentanglement. On the other hand, as we increase the number of latent dimensions (top-down direction in the figure), reconstruction quality for fine details is enhanced.
• Figure 4: M.A.D values for 128-d network on the top and 8-d network on the bottom. From left to right the disentanglement weight ($\lambda$) is increased. Too low weights result in leak of information among different dimensions, while too high values cause loss of details which causes better disentanglement, yet less useful features. Interestingly the 128-d and 8-d networks agree on the number of informative features at $\lambda=0.3$.
• Figure 5: Scatter plots illustrating mutual behaviour of pairs of latent dimensions. On the top, there is little to no significant correlation between the two. In contrast, the bottom two plots show clear correlation between exemplar dimension pairs, in networks where $\lambda$ has been too low, which is a strong hint for failure of disentanglement. In such cases, a high M.A.D does not directly translate to possession of exclusive information. Contrary to intuition, the less structured the plots are, the more successful the disentanglement has been. Different colors show different spectral classes and are used for illustration purposes only.