The adoption of Vision Transformers (ViTs) in resource-constrained applications necessitates improvements in inference throughput. To this end several token pruning and merging approaches have been proposed that improve efficiency by successively reducing the number of tokens. However, it remains an open problem to design a token reduction method that is fast, maintains high performance, and is applicable to various vision tasks. In this work, we present a token pruner that uses auxiliary prediction heads that learn to select tokens end-to-end based on task relevance. These auxiliary heads can be removed after training, leading to throughput close to that of a random pruner. We evaluate our method on image classification, semantic segmentation, object detection, and instance segmentation, and show speedups of 1.5 to 4x with small drops in performance. As a best case, on the ADE20k semantic segmentation benchmark, we observe a 2x speedup relative to the no-pruning baseline, with a negligible performance penalty of 0.1 median mIoU across 5 seeds.

Conformal Prediction offers a powerful framework for quantifying uncertainty in machine learning models, enabling the construction of prediction sets with finite-sample validity guarantees. While easily adaptable to non-probabilistic models, applying conformal prediction to probabilistic generative models, such as Normalising Flows is not straightforward. This work proposes a novel method to conformalise conditional normalising flows, specifically addressing the problem of obtaining prediction regions for multi-step time series forecasting. Our approach leverages the flexibility of normalising flows to generate potentially disjoint prediction regions, leading to improved predictive efficiency in the presence of potential multimodal predictive distributions.

As existing CITs are limited in their applicability to complex, high-dimensional variables such as images, we introduce deep nonparametric CITs (DNCITs). The DNCITs combine embedding maps, which extract feature representations of high-dimensional variables, with nonparametric CITs applicable to these feature representations. For the embedding maps, we derive general properties on their parameter estimators to obtain valid DNCITs and show that these properties include embedding maps learned through (conditional) unsupervised or transfer learning. For the nonparametric CITs, appropriate tests are selected and adapted to be applicable to feature representations. Through simulations, we investigate the performance of the DNCITs for different embedding maps and nonparametric CITs under varying confounder dimensions and confounder relationships. We apply the DNCITs to brain MRI scans and behavioral traits, given confounders, of healthy individuals from the UK Biobank (UKB), confirming null results from a number of ambiguous personality neuroscience studies with a larger data set and with our more powerful tests. In addition, in a confounder control study, we apply the DNCITs to brain MRI scans and a confounder set to test for sufficient confounder control, leading to a potential reduction in the confounder dimension under improved confounder control compared to existing state-of-the-art confounder control studies for the UKB. Finally, we provide an R package implementing the DNCITs.

The field of deep generative modeling has grown rapidly and consistently over the years. With the availability of massive amounts of training data coupled with advances in scalable unsupervised learning paradigms, recent large-scale generative models show tremendous promise in synthesizing high-resolution images and text, as well as structured data such as videos and molecules. However, we argue that current large-scale generative AI models do not sufficiently address several fundamental issues that hinder their widespread adoption across domains. In this work, we aim to identify key unresolved challenges in modern generative AI paradigms that should be tackled to further enhance their capabilities, versatility, and reliability. By identifying these challenges, we aim to provide researchers with valuable insights for exploring fruitful research directions, thereby fostering the development of more robust and accessible generative AI solutions.

Normalising flows are statistical models that transform a complex density into a simpler density through the use of bijective transformations enabling both density estimation and data generation from a single model. In the context of image modelling, the predominant choice has been the Glow-based architecture, whereas alternative architectures remain largely unexplored in the research community. In this work, we propose a novel architecture called MixerFlow, based on the MLP-Mixer architecture, further unifying the generative and discriminative modelling architectures. MixerFlow offers an effective mechanism for weight sharing for flow-based models. Our results demonstrate better density estimation on image datasets under a fixed computational budget and scales well as the image resolution increases, making MixeFlow a powerful yet simple alternative to the Glow-based architectures. We also show that MixerFlow provides more informative embeddings than Glow-based architectures.

Normalising Flows are non-parametric statistical models characterised by their dual capabilities of density estimation and generation. This duality requires an inherently invertible architecture. However, the requirement of invertibility imposes constraints on their expressiveness, necessitating a large number of parameters and innovative architectural designs to achieve good results. Whilst flow-based models predominantly rely on neural-network-based transformations for expressive designs, alternative transformation methods have received limited attention. In this work, we present Ferumal flow, a novel kernelised normalising flow paradigm that integrates kernels into the framework. Our results demonstrate that a kernelised flow can yield competitive or superior results compared to neural network-based flows whilst maintaining parameter efficiency. Kernelised flows excel especially in the low-data regime, enabling flexible non-parametric density estimation in applications with sparse data availability.

In this paper we present a practical Bayesian self-supervised learning method with Cyclical Stochastic Gradient Hamiltonian Monte Carlo (cSGHMC). Within this framework, we place a prior over the parameters of a self-supervised learning model and use cSGHMC to approximate the high dimensional and multimodal posterior distribution over the embeddings. By exploring an expressive posterior over the embeddings, Bayesian self-supervised learning produces interpretable and diverse representations. Marginalizing over these representations yields a significant gain in performance, calibration and out-of-distribution detection on a variety of downstream classification tasks. We provide experimental results on multiple classification tasks on four challenging datasets. Moreover, we demonstrate the effectiveness of the proposed method in out-of-distribution detection using the SVHN and CIFAR-10 datasets.

We consider the problem of identifying the signal shared between two one-dimensional target variables, in the presence of additional multivariate observations. Canonical Correlation Analysis (CCA)-based methods have traditionally been used to identify shared variables, however, they were designed for multivariate targets and only offer trivial solutions for univariate cases. In the context of Multi-Task Learning (MTL), various models were postulated to learn features that are sparse and shared across multiple tasks. However, these methods were typically evaluated by their predictive performance. To the best of our knowledge, no prior studies systematically evaluated models in terms of correctly recovering the shared signal. Here, we formalize the setting of univariate shared information retrieval, and propose ICM, an evaluation metric which can be used in the presence of ground-truth labels, quantifying 3 aspects of the learned shared features. We further propose Deep Canonical Information Decomposition (DCID) - a simple, yet effective approach for learning the shared variables. We benchmark the models on a range of scenarios on synthetic data with known ground-truths and observe DCID outperforming the baselines in a wide range of settings. Finally, we demonstrate a real-life application of DCID on brain Magnetic Resonance Imaging (MRI) data, where we are able to extract more accurate predictors of changes in brain regions and obesity.

Normalizing flows are powerful non-parametric statistical models that function as a hybrid between density estimators and generative models. Current learning algorithms for normalizing flows assume that data points are sampled independently, an assumption that is frequently violated in practice, which may lead to erroneous density estimation and data generation. We propose a likelihood objective of normalizing flows incorporating dependencies between the data points, for which we derive a flexible and efficient learning algorithm suitable for different dependency structures. We show that respecting dependencies between observations can improve empirical results on both synthetic and real-world data, and leads to higher statistical power in a downstream application to genome-wide association studies.

High-resolution images are prevalent in various applications, such as autonomous driving and computer-aided diagnosis. However, training neural networks on such images is computationally challenging and easily leads to out-of-memory errors even on modern GPUs. We propose a simple method, Iterative Patch Selection (IPS), which decouples the memory usage from the input size and thus enables the processing of arbitrarily large images under tight hardware constraints. IPS achieves this by selecting only the most salient patches, which are then aggregated into a global representation for image recognition. For both patch selection and aggregation, a cross-attention based transformer is introduced, which exhibits a close connection to Multiple Instance Learning. Our method demonstrates strong performance and has wide applicability across different domains, training regimes and image sizes while using minimal accelerator memory. For example, we are able to finetune our model on whole-slide images consisting of up to 250k patches (>16 gigapixels) with only 5 GB of GPU VRAM at a batch size of 16. Image recognition has made great strides in recent years, spawning landmark architectures such as AlexNet (Krizhevsky et al., 2012) or ResNet (He et al., 2016). These networks are typically designed and optimized for datasets like ImageNet (Russakovsky et al., 2015), which consist of natural images well below one megapixel. In contrast, realworld applications often rely on high-resolution images that reveal detailed information about an object of interest. For example, in self-driving cars, megapixel images are beneficial to recognize distant traffic signs far in advance and react in time (Sahin, 2019). In medical imaging, a pathology diagnosis system has to process gigapixel microscope slides to recognize cancer cells, as illustrated in Figure 1.