As deep generative models have progressed, recent work has shown them to be capable of memorizing and reproducing training datapoints when deployed. These findings call into question the usability of generative models, especially in light of the legal and privacy risks brought about by memorization. To better understand this phenomenon, we propose the manifold memorization hypothesis (MMH), a geometric framework which leverages the manifold hypothesis into a clear language in which to reason about memorization. We propose to analyze memorization in terms of the relationship between the dimensionalities of $(i)$ the ground truth data manifold and $(ii)$ the manifold learned by the model. This framework provides a formal standard for "how memorized" a datapoint is and systematically categorizes memorized data into two types: memorization driven by overfitting and memorization driven by the underlying data distribution. By analyzing prior work in the context of the MMH, we explain and unify assorted observations in the literature. We empirically validate the MMH using synthetic data and image datasets up to the scale of Stable Diffusion, developing new tools for detecting and preventing generation of memorized samples in the process.

Large-scale vision models have become integral in many applications due to their unprecedented performance and versatility across downstream tasks. However, the robustness of these foundation models has primarily been explored for a single task, namely image classification. The vulnerability of other common vision tasks, such as semantic segmentation and depth estimation, remains largely unknown. We present a comprehensive empirical evaluation of the adversarial robustness of self-supervised vision encoders across multiple downstream tasks. Our attacks operate in the encoder embedding space and at the downstream task output level. In both cases, current state-of-the-art adversarial fine-tuning techniques tested only for classification significantly degrade clean and robust performance on other tasks. Since the purpose of a foundation model is to cater to multiple applications at once, our findings reveal the need to enhance encoder robustness more broadly. Our code is available at ${github.com/layer6ai-labs/ssl-robustness}$.

Advances in deep generative modelling have not translated well to tabular data. We argue that this is caused by a mismatch in structure between popular generative models and discriminative models of tabular data. We thus devise a technique to turn TabPFN -- a highly performant transformer initially designed for in-context discriminative tabular tasks -- into an energy-based generative model, which we dub TabPFGen. This novel framework leverages the pre-trained TabPFN as part of the energy function and does not require any additional training or hyperparameter tuning, thus inheriting TabPFN's in-context learning capability. We can sample from TabPFGen analogously to other energy-based models. We demonstrate strong results on standard generative modelling tasks, including data augmentation, class-balancing, and imputation, unlocking a new frontier of tabular data generation.

We systematically study a wide variety of generative models spanning semantically-diverse image datasets to understand and improve the feature extractors and metrics used to evaluate them. Using best practices in psychophysics, we measure human perception of image realism for generated samples by conducting the largest experiment evaluating generative models to date, and find that no existing metric strongly correlates with human evaluations. Comparing to 17 modern metrics for evaluating the overall performance, fidelity, diversity, rarity, and memorization of generative models, we find that the state-of-the-art perceptual realism of diffusion models as judged by humans is not reflected in commonly reported metrics such as FID. This discrepancy is not explained by diversity in generated samples, though one cause is over-reliance on Inception-V3. We address these flaws through a study of alternative self-supervised feature extractors, find that the semantic information encoded by individual networks strongly depends on their training procedure, and show that DINOv2-ViT-L/14 allows for much richer evaluation of generative models. Next, we investigate data memorization, and find that generative models do memorize training examples on simple, smaller datasets like CIFAR10, but not necessarily on more complex datasets like ImageNet. However, our experiments show that current metrics do not properly detect memorization: none in the literature is able to separate memorization from other phenomena such as underfitting or mode shrinkage. To facilitate further development of generative models and their evaluation we release all generated image datasets, human evaluation data, and a modular library to compute 17 common metrics for 9 different encoders at https://github.com/layer6ai-labs/dgm-eval.

Self-supervised representation learning~(SSRL) has advanced considerably by exploiting the transformation invariance assumption under artificially designed data augmentations. While augmentation-based SSRL algorithms push the boundaries of performance in computer vision and natural language processing, they are often not directly applicable to other data modalities, and can conflict with application-specific data augmentation constraints. This paper presents an SSRL approach that can be applied to any data modality and network architecture because it does not rely on augmentations or masking. Specifically, we show that high-quality data representations can be learned by reconstructing random data projections. We evaluate the proposed approach on a wide range of representation learning tasks that span diverse modalities and real-world applications. We show that it outperforms multiple state-of-the-art SSRL baselines. Due to its wide applicability and strong empirical results, we argue that learning from randomness is a fruitful research direction worthy of attention and further study.

We use a contrastive self-supervised learning framework to estimate distances to galaxies from their photometric images. We incorporate data augmentations from computer vision as well as an application-specific augmentation accounting for galactic dust. We find that the resulting visual representations of galaxy images are semantically useful and allow for fast similarity searches, and can be successfully fine-tuned for the task of redshift estimation. We show that (1) pretraining on a large corpus of unlabeled data followed by fine-tuning on some labels can attain the accuracy of a fully-supervised model which requires 2-4x more labeled data, and (2) that by fine-tuning our self-supervised representations using all available data labels in the Main Galaxy Sample of the Sloan Digital Sky Survey (SDSS), we outperform the state-of-the-art supervised learning method.

Submitted to The Astrophysical Journal Letters ABSTRACT Sky surveys are the largest data generators in astronomy, making automated tools for extracting meaningful scientific information an absolute necessity. We show that, without the need for labels, self-supervised learning recovers representations of sky survey images that are semantically useful for a variety of scientific tasks. These representations can be directly used as features, or fine-tuned, to outperform supervised methods trained only on labeled data. We apply a contrastive learning framework on multi-band galaxy photometry from the Sloan Digital Sky Survey (SDSS), to learn image representations. We then use them for galaxy morphology classification, and fine-tune them for photometric redshift estimation, using labels from the Galaxy Zoo 2 dataset and SDSS spectroscopy. In both downstream tasks, using the same learned representations, we outperform the supervised stateof-the-art results, and we show that our approach can achieve the accuracy of supervised models while using 2-4 times fewer labels for training. INTRODUCTION the quantity and quality of (manually assigned) image labels. Observing and imaging objects in the sky has been Serendipitous discovery of an ionization echo from a the main driver of the scientific discovery process in astronomy, recently faded quasar (Lintott et al. 2009), and the cumbersome because doing controlled experiments is not a search for similar systems that followed (Keel viable option. It in the 1990s, spearheaded by SDSS (Gunn et al. 1998, demonstrates the need for methods which allow for the 2006), has rendered obsolete the approach of manual discovery of truly unusual and previously unseen objects, inspection of images by an expert.