Steering methods manipulate the representations of large language models (LLMs) to induce responses that have desired properties, e.g., truthfulness, offering a promising approach for LLM alignment without the need for fine-tuning. Traditionally, steering has relied on supervision, such as from contrastive pairs of prompts that vary in a single target concept, which is costly to obtain and limits the speed of steering research. An appealing alternative is to use unsupervised approaches such as sparse autoencoders (SAEs) to map LLM embeddings to sparse representations that capture human-interpretable concepts. However, without further assumptions, SAEs may not be identifiable: they could learn latent dimensions that entangle multiple concepts, leading to unintentional steering of unrelated properties. We introduce Sparse Shift Autoencoders (SSAEs) that instead map the differences between embeddings to sparse representations. Crucially, we show that SSAEs are identifiable from paired observations that vary in \textit{multiple unknown concepts}, leading to accurate steering of single concepts without the need for supervision. We empirically demonstrate accurate steering across semi-synthetic and real-world language datasets using Llama-3.1 embeddings.

Generative modeling of single-cell RNA-seq data has shown invaluable potential in community-driven tasks such as trajectory inference, batch effect removal and gene expression generation. However, most recent deep models generating synthetic single cells from noise operate on pre-processed continuous gene expression approximations, ignoring the inherently discrete and over-dispersed nature of single-cell data, which limits downstream applications and hinders the incorporation of robust noise models. Moreover, crucial aspects of deep-learning-based synthetic single-cell generation remain underexplored, such as controllable multi-modal and multi-label generation and its role in the performance enhancement of downstream tasks. This work presents Cell Flow for Generation (CFGen), a flow-based conditional generative model for multi-modal single-cell counts, which explicitly accounts for the discrete nature of the data. Our results suggest improved recovery of crucial biological data characteristics while accounting for novel generative tasks such as conditioning on multiple attributes and boosting rare cell type classification via data augmentation. By showcasing CFGen on a diverse set of biological datasets and settings, we provide evidence of its value to the fields of computational biology and deep generative models.

Causal reasoning can be considered a cornerstone of intelligent systems. Having access to an underlying causal graph comes with the promise of cause-effect estimation and the identification of efficient and safe interventions. However, learning causal representations remains a major challenge, due to the complexity of many real-world systems. Previous works on causal representation learning have mostly focused on Variational Auto-Encoders (VAE). These methods only provide representations from a point estimate, and they are unsuitable to handle high dimensions. To overcome these problems, we proposed a new Diffusion-based Causal Representation Learning (DCRL) algorithm. This algorithm uses diffusion-based representations for causal discovery. DCRL offers access to infinite dimensional latent codes, which encode different levels of information in the latent code. In a first proof of principle, we investigate the use of DCRL for causal representation learning. We further demonstrate experimentally that this approach performs comparably well in identifying the causal structure and causal variables.

Diffusion models may be viewed as hierarchical variational autoencoders (VAEs) with two improvements: parameter sharing for the conditional distributions in the generative process and efficient computation of the loss as independent terms over the hierarchy. We consider two changes to the diffusion model that retain these advantages while adding flexibility to the model. Firstly, we introduce a data- and depth-dependent mean function in the diffusion process, which leads to a modified diffusion loss. Our proposed framework, DiffEnc, achieves state-of-the-art likelihood on CIFAR-10. Secondly, we let the ratio of the noise variance of the reverse encoder process and the generative process be a free weight parameter rather than being fixed to 1. This leads to theoretical insights: For a finite depth hierarchy, the evidence lower bound (ELBO) can be used as an objective for a weighted diffusion loss approach and for optimizing the noise schedule specifically for inference. For the infinite-depth hierarchy, on the other hand, the weight parameter has to be 1 to have a well-defined ELBO.

Despite tremendous progress over the past decade, deep learning methods generally fall short of human-level systematic generalization. It has been argued that explicitly capturing the underlying structure of data should allow connectionist systems to generalize in a more predictable and systematic manner. Indeed, evidence in humans suggests that interpreting the world in terms of symbol-like compositional entities may be crucial for intelligent behavior and high-level reasoning. Another common limitation of deep learning systems is that they require large amounts of training data, which can be expensive to obtain. In representation learning, large datasets are leveraged to learn generic data representations that may be useful for efficient learning of arbitrary downstream tasks. This thesis is about structured representation learning. We study methods that learn, with little or no supervision, representations of unstructured data that capture its hidden structure. In the first part of the thesis, we focus on representations that disentangle the explanatory factors of variation of the data. We scale up disentangled representation learning to a novel robotic dataset, and perform a systematic large-scale study on the role of pretrained representations for out-of-distribution generalization in downstream robotic tasks. The second part of this thesis focuses on object-centric representations, which capture the compositional structure of the input in terms of symbol-like entities, such as objects in visual scenes. Object-centric learning methods learn to form meaningful entities from unstructured input, enabling symbolic information processing on a connectionist substrate. In this study, we train a selection of methods on several common datasets, and investigate their usefulness for downstream tasks and their ability to generalize out of distribution.

In representation learning, a common approach is to seek representations which disentangle the underlying factors of variation. Eastwood & Williams (2018) proposed three metrics for quantifying the quality of such disentangled representations: disentanglement (D), completeness (C) and informativeness (I). In this work, we first connect this DCI framework to two common notions of linear and nonlinear identifiability, thereby establishing a formal link between disentanglement and the closely-related field of independent component analysis. We then propose an extended DCI-ES framework with two new measures of representation quality - explicitness (E) and size (S) - and point out how D and C can be computed for black-box predictors. Our main idea is that the functional capacity required to use a representation is an important but thus-far neglected aspect of representation quality, which we quantify using explicitness or ease-of-use (E). We illustrate the relevance of our extensions on the MPI3D and Cars3D datasets.

The ability to assess the robustness of image classifiers to a diverse set of manipulations is essential to their deployment in the real world. Recently, semantic manipulations of real images have been considered for this purpose, as they may not arise using standard adversarial settings. However, such semantic manipulations are often limited to style, color or attribute changes. While expressive, these manipulations do not consider the full capacity of a pretrained generator to affect adversarial image manipulations. In this work, we aim at leveraging the full capacity of a pretrained image generator to generate highly detailed, diverse and photorealistic image manipulations. Inspired by recent GAN-based image inversion methods, we propose a method called Adversarial Pivotal Tuning (APT). APT first finds a pivot latent space input to a pretrained generator that best reconstructs an input image. It then adjusts the weights of the generator to create small, but semantic, manipulations which fool a pretrained classifier. Crucially, APT changes both the input and the weights of the pretrained generator, while preserving its expressive latent editing capability, thus allowing the use of its full capacity in creating semantic adversarial manipulations. We demonstrate that APT generates a variety of semantic image manipulations, which preserve the input image class, but which fool a variety of pretrained classifiers. We further demonstrate that classifiers trained to be robust to other robustness benchmarks, are not robust to our generated manipulations and propose an approach to improve the robustness towards our generated manipulations. Code available at: https://captaine.github.io/apt/

Disentanglement is hypothesized to be beneficial towards a number of downstream tasks. However, a common assumption in learning disentangled representations is that the data generative factors are statistically independent. As current methods are almost solely evaluated on toy datasets where this ideal assumption holds, we investigate their performance in hierarchical settings, a relevant feature of real-world data. In this work, we introduce Boxhead, a dataset with hierarchically structured ground-truth generative factors. We use this novel dataset to evaluate the performance of state-of-the-art autoencoder-based disentanglement models and observe that hierarchical models generally outperform single-layer VAEs in terms of disentanglement of hierarchically arranged factors.

Learning data representations that are useful for various downstream tasks is a cornerstone of artificial intelligence. While existing methods are typically evaluated on downstream tasks such as classification or generative image quality, we propose to assess representations through their usefulness in downstream control tasks, such as reaching or pushing objects. By training over 10,000 reinforcement learning policies, we extensively evaluate to what extent different representation properties affect out-of-distribution (OOD) generalization. Finally, we demonstrate zero-shot transfer of these policies from simulation to the real world, without any domain randomization or fine-tuning. This paper aims to establish the first systematic characterization of the usefulness of learned representations for real-world OOD downstream tasks.

The idea behind object-centric representation learning is that natural scenes can better be modeled as compositions of objects and their relations as opposed to distributed representations. This inductive bias can be injected into neural networks to potentially improve systematic generalization and learning efficiency of downstream tasks in scenes with multiple objects. In this paper, we train state-of-the-art unsupervised models on five common multi-object datasets and evaluate segmentation accuracy and downstream object property prediction. In addition, we study systematic generalization and robustness by investigating the settings where either single objects are out-of-distribution -- e.g., having unseen colors, textures, and shapes -- or global properties of the scene are altered -- e.g., by occlusions, cropping, or increasing the number of objects. From our experimental study, we find object-centric representations to be generally useful for downstream tasks and robust to shifts in the data distribution, especially if shifts affect single objects.