Decomposing multimodal embedding spaces with group-sparse autoencoders

TL;DR

The paper tackles the interpretability of multimodal embeddings by showing that standard sparse autoencoders trained on aligned multimodal spaces tend to learn modality-split dictionaries, hindering cross-modal alignment. It introduces Masked Group-Sparse Autoencoders (MGSAE) that use a group-sparse loss and cross-modal random masking to promote truly multimodal dictionaries and shared sparse representations. The authors formalize a multimodal monosemanticity metric (MMS) and prove that a split dictionary implies a non-split, better-aligned dictionary under mild conditions. Empirical results on CLIP (image/text) and CLAP (music/text) demonstrate that MGSAE learns more multimodal concepts, reduces dead neurons, improves zero-shot cross-modal tasks, and enhances interpretability via concept naming. This work advances controllable, interpretable multimodal representations with potential extensions to more modalities and unpaired data.

Abstract

The Linear Representation Hypothesis asserts that the embeddings learned by neural networks can be understood as linear combinations of features corresponding to high-level concepts. Based on this ansatz, sparse autoencoders (SAEs) have recently become a popular method for decomposing embeddings into a sparse combination of linear directions, which have been shown empirically to often correspond to human-interpretable semantics. However, recent attempts to apply SAEs to multimodal embedding spaces (such as the popular CLIP embeddings for image/text data) have found that SAEs often learn "split dictionaries", where most of the learned sparse features are essentially unimodal, active only for data of a single modality. In this work, we study how to effectively adapt SAEs for the setting of multimodal embeddings while ensuring multimodal alignment. We first argue that the existence of a split dictionary decomposition on an aligned embedding space implies the existence of a non-split dictionary with improved modality alignment. Then, we propose a new SAE-based approach to multimodal embedding decomposition using cross-modal random masking and group-sparse regularization. We apply our method to popular embeddings for image/text (CLIP) and audio/text (CLAP) data and show that, compared to standard SAEs, our approach learns a more multimodal dictionary while reducing the number of dead neurons and improving feature semanticity. We finally demonstrate how this improvement in alignment of concepts between modalities can enable improvements in the interpretability and control of cross-modal tasks.

Figures (7)

• Figure 1: Split versus multimodal dictionaries: (Left) Standard SAEs trained on aligned embeddings from different modalities (like CLIP) tend to learn dictionary vectors ("concepts") which activate only for input embeddings of one modality. (Right) We develop an SAE-based approach for learning dictionaries where multimodal embeddings containing the same semantic information have similar sparse dictionary decompositions.
• Figure 2: Masked group-sparse autoencoder for multimodal concept extraction: Our proposed approach is trained on paired data. Embeddings are encoded to a higher-dimensional vector, masked (with the same mask for each modality), sparsified using $\text{TopK}$, and decoded with a linear layer.
• Figure 3: Number of neurons activating for each individual modality, both modalities, and neither modality. Left: models trained on CLIP embeddings, validation on CC3M val. set. Right: models trained on CLAP embeddings, validation on MusicBench.
• Figure 4: Multimodal monosemanticity (MMS) scores for each feature, arranged in descending order (higher is better). Neuron index (on the horizontal axis) is normalized due the difference in dimensionality between the "No SAE" case ($d=512$) and the SAE variants ($d=16\cdot512$). Top row: MMS scores for models trained on CLIP embeddings for image/text data. Bottom row: MMS scores for models trained on CLAP embeddings for music/text data.
• Figure 5: Leading concepts contributing to a classification of "blonde" on the CelebA dataset, extracted using a standard SAE (left) and MGSAE (right).

Theorems & Definitions (3)

• Definition 1: Modality-split dictionary
• Theorem 1
• proof