Gemma E. Moran

Gerasimos Chatzoudis, Konstantinos D. Polyzos, Zhuowei Li et al.

Understanding the internal activations of Vision Transformers (ViTs) is critical for building interpretable and trustworthy models. While Sparse Autoencoders (SAEs) have been used to extract human-interpretable features, they operate on individual layers and fail to capture the cross-layer computational structure of Transformers, as well as the relative significance of each layer in forming the last-layer representation. Alternatively, we introduce the adoption of Cross-Layer Transcoders (CLTs) as reliable, sparse, and depth-aware proxy models for MLP blocks in ViTs. CLTs use an encoder-decoder scheme to reconstruct each post-MLP activation from learned sparse embeddings of preceding layers, yielding a linear decomposition that transforms the final representation of ViTs from an opaque embedding into an additive, layer-resolved construction that enables faithful attribution and process-level interpretability. We train CLTs on CLIP ViT-B/32 and ViT-B/16 across CIFAR-100, COCO, and ImageNet-100. We show that CLTs achieve high reconstruction fidelity with post-MLP activations while preserving and even improving, in some cases, CLIP zero-shot classification accuracy. In terms of interpretability, we show that the cross-layer contribution scores provide faithful attribution, revealing that the final representation is concentrated in a smaller set of dominant layer-wise terms whose removal degrades performance and whose retention largely preserves it. These results showcase the significance of adopting CLTs as an alternative interpretable proxy of ViTs in the vision domain.

Gemma E. Moran, Bryon Aragam

Recent developments in generative artificial intelligence (AI) rely on machine learning techniques such as deep learning and generative modeling to achieve state-of-the-art performance across wide-ranging domains. These methods' surprising performance is due in part to their ability to learn implicit "representations'' of complex, multi-modal data. Unfortunately, deep neural networks are notoriously black boxes that obscure these representations, making them difficult to interpret or analyze. To resolve these difficulties, one approach is to build new interpretable neural network models from the ground up. This is the goal of the emerging field of causal representation learning (CRL) that uses causality as a vector for building flexible, interpretable, and transferable generative AI. CRL can be seen as a culmination of three intrinsically statistical problems: (i) latent variable models such as factor analysis; (ii) causal graphical models with latent variables; and (iii) nonparametric statistics and deep learning. This paper reviews recent progress in CRL from a statistical perspective, focusing on connections to classical models and statistical and causal identifiablity results. This review also highlights key application areas, implementation strategies, and open statistical questions in CRL.

Gemma E. Moran, Anandi Krishnan

High-dimensional data often exhibit variation that can be captured by lower dimensional factors. For high-dimensional data from multiple studies or environments, one goal is to understand which underlying factors are common to all studies, and which factors are study or environment-specific. As a particular example, we consider platelet gene expression data from patients in different disease groups. In this data, factors correspond to clusters of genes which are co-expressed; we may expect some clusters (or biological pathways) to be active for all diseases, while some clusters are only active for a specific disease. To learn these factors, we consider a nonlinear multi-study factor model, which allows for both shared and specific factors. To fit this model, we propose a multi-study sparse variational autoencoder. The underlying model is sparse in that each observed feature (i.e. each dimension of the data) depends on a small subset of the latent factors. In the genomics example, this means each gene is active in only a few biological processes. Further, the model implicitly induces a penalty on the number of latent factors, which helps separate the shared factors from the group-specific factors. We prove that the latent factors are identified, and demonstrate our method recovers meaningful factors in the platelet gene expression data.

Gerasimos Chatzoudis, Zhuowei Li, Gemma E. Moran et al.

Steering vision foundation models at inference time without retraining or access to large labeled datasets is a desirable yet challenging objective, particularly in dynamic or resource-constrained settings. In this paper, we introduce Visual Sparse Steering (VS2), a lightweight, test-time method that guides vision models using steering vectors derived from sparse features learned by top-$k$ Sparse Autoencoders without requiring contrastive data. Specifically, VS2 surpasses zero-shot CLIP by 4.12% on CIFAR-100, 1.08% on CUB-200, and 1.84% on Tiny-ImageNet. We further propose VS2++, a retrieval-augmented variant that selectively amplifies relevant sparse features using pseudo-labeled neighbors at inference time. With oracle positive/negative sets, VS2++ achieves absolute top-1 gains over CLIP zero-shot of up to 21.44% on CIFAR-100, 7.08% on CUB-200, and 20.47% on Tiny-ImageNet. Interestingly, VS2 and VS2++ raise per-class accuracy by up to 25% and 38%, respectively, showing that sparse steering benefits specific classes by disambiguating visually or taxonomically proximate categories rather than providing a uniform boost. Finally, to better align the sparse features learned through the SAE reconstruction task with those relevant for downstream performance, we propose Prototype-Aligned Sparse Steering (PASS). By incorporating a prototype-alignment loss during SAE training, using labels only during training while remaining fully test-time unsupervised, PASS consistently, though modestly, outperforms VS2, achieving a 6.12% gain over VS2 only on CIFAR-100 with ViT-B/32.

Gemma E. Moran, Dhanya Sridhar, Yixin Wang et al.

We develop the sparse VAE for unsupervised representation learning on high-dimensional data. The sparse VAE learns a set of latent factors (representations) which summarize the associations in the observed data features. The underlying model is sparse in that each observed feature (i.e. each dimension of the data) depends on a small subset of the latent factors. As examples, in ratings data each movie is only described by a few genres; in text data each word is only applicable to a few topics; in genomics, each gene is active in only a few biological processes. We prove such sparse deep generative models are identifiable: with infinite data, the true model parameters can be learned. (In contrast, most deep generative models are not identifiable.) We empirically study the sparse VAE with both simulated and real data. We find that it recovers meaningful latent factors and has smaller heldout reconstruction error than related methods.

Gemma E. Moran, David M. Blei, Rajesh Ranganath

Bayesian modeling helps applied researchers articulate assumptions about their data and develop models tailored for specific applications. Thanks to good methods for approximate posterior inference, researchers can now easily build, use, and revise complicated Bayesian models for large and rich data. These capabilities, however, bring into focus the problem of model criticism. Researchers need tools to diagnose the fitness of their models, to understand where they fall short, and to guide their revision. In this paper we develop a new method for Bayesian model criticism, the population predictive check (Pop-PC). Pop-PCs are built on posterior predictive checks (PPCs), a seminal method that checks a model by assessing the posterior predictive distribution on the observed data. However, PPCs use the data twice -- both to calculate the posterior predictive and to evaluate it -- which can lead to overconfident assessments of the quality of a model. Pop-PCs, in contrast, compare the posterior predictive distribution to a draw from the population distribution, a heldout dataset. This method blends Bayesian modeling with frequenting assessment. Unlike the PPC, we prove that the Pop-PC is properly calibrated. Empirically, we study Pop-PC on classical regression and a hierarchical model of text data.