Re-envisioning Euclid Galaxy Morphology: Identifying and Interpreting Features with Sparse Autoencoders

TL;DR

The paper addresses interpreting deep morphology representations and discovering rare galaxy features in large surveys by applying Matryoshka sparse autoencoders (SAEs) to embeddings from supervised Zoobot and self-supervised MAE models on Euclid Q1 data. It demonstrates that SAEs yield monosemantic, interpretable features that align with Galaxy Zoo classifications better than PCA, while also revealing features beyond the traditional taxonomy; MAE-based embeddings achieve near-superhuman reconstruction, underscoring the richness of information captured by self-supervised representations. The work shows SAEs as scalable discovery engines that can surface novel or artifact-related phenomena in upcoming surveys, though interpretability and artifact sensitivity remain challenges. Public release of the MAE model and SAE code enables community-driven exploration of morphology concepts across next-generation astronomical datasets.

Abstract

Sparse Autoencoders (SAEs) can efficiently identify candidate monosemantic features from pretrained neural networks for galaxy morphology. We demonstrate this on Euclid Q1 images using both supervised (Zoobot) and new self-supervised (MAE) models. Our publicly released MAE achieves superhuman image reconstruction performance. While a Principal Component Analysis (PCA) on the supervised model primarily identifies features already aligned with the Galaxy Zoo decision tree, SAEs can identify interpretable features outside of this framework. SAE features also show stronger alignment than PCA with Galaxy Zoo labels. Although challenges in interpretability remain, SAEs provide a powerful engine for discovering astrophysical phenomena beyond the confines of human-defined classification.

Figures (5)

• Figure 1: True images (top row), masked inputs (middle row), and reconstructed outputs (bottom row), grouped as follows. Top: challenging images with successful MAE reconstructions. Lower left: images with missing local details (e.g., small starforming clumps, multiple background sources with marginal separation). These cannot be reconstructed if not at least partially included in an unmasked patch. Masking is only applied during training and so these details are still in the embeddings. Lower right: random images (first five) demonstrating that near-perfect reconstructions are the norm.
• Figure 2: Top 10 examples for first five features extracted via PCA (upper), and SAEs (lower), for embeddings extracted from supervised (left) and self-supervised (right) models.
• Figure 3: Top SAE features (red) are generally more aligned with GZ morphological classifications than PCA (blue) for both the supervised model embeddings (left) and self-supervised model embeddings (right). Both panels show the distribution of maximum Spearman rank correlation $r$ values between top 64 features and any GZ class.
• Figure 4: How aligned are features with GZ, just like in Figure \ref{['fig:feature_alignment']}? We show results for both supervised (left) and self-supervised (right) approaches. This time we use all of the SAE activations, binned by their Matryoshka group sizes (progressing from lighter to darker red). In all panels, the top 64 elements from PCA are shown in a thin dashed blue line.
• Figure 5: How unpredictable---or perhaps, "novel"---are our features via linear combinations of GZ vote fractions? Figure in same style as Figure \ref{['fig:appendix_feature_alignment']}.