Sebastian Mair

Dominik Fay, Sebastian Mair, Jens Sjölund

For scalable machine learning on large data sets, subsampling a representative subset is a common approach for efficient model training. This is often achieved through importance sampling, whereby informative data points are sampled more frequently. In this paper, we examine the privacy properties of importance sampling, focusing on an individualized privacy analysis. We find that, in importance sampling, privacy is well aligned with utility but at odds with sample size. Based on this insight, we propose two approaches for constructing sampling distributions: one that optimizes the privacy-efficiency trade-off; and one based on a utility guarantee in the form of coresets. We evaluate both approaches empirically in terms of privacy, efficiency, and accuracy on the differentially private $k$-means problem. We observe that both approaches yield similar outcomes and consistently outperform uniform sampling across a wide range of data sets. Our code is available on GitHub: https://github.com/smair/personalized-privacy-amplification-via-importance-sampling

Sebastian Mair, Jens Sjölund

Archetypal analysis is a matrix factorization method with convexity constraints. Due to local minima, a good initialization is essential, but frequently used initialization methods yield either sub-optimal starting points or are prone to get stuck in poor local minima. In this paper, we propose archetypal analysis++ (AA++), a probabilistic initialization strategy for archetypal analysis that sequentially samples points based on their influence on the objective function, similar to $k$-means++. In fact, we argue that $k$-means++ already approximates the proposed initialization method. Furthermore, we suggest to adapt an efficient Monte Carlo approximation of $k$-means++ to AA++. In an extensive empirical evaluation of 15 real-world data sets of varying sizes and dimensionalities and considering two pre-processing strategies, we show that AA++ almost always outperforms all baselines, including the most frequently used ones.

Zheng Zhao, Sebastian Mair, Thomas B. Schön et al.

Recently, partial Bayesian neural networks (pBNNs), which only consider a subset of the parameters to be stochastic, were shown to perform competitively with full Bayesian neural networks. However, pBNNs are often multi-modal in the latent variable space and thus challenging to approximate with parametric models. To address this problem, we propose an efficient sampling-based training strategy, wherein the training of a pBNN is formulated as simulating a Feynman--Kac model. We then describe variations of sequential Monte Carlo samplers that allow us to simultaneously estimate the parameters and the latent posterior distribution of this model at a tractable computational cost. Using various synthetic and real-world datasets we show that our proposed training scheme outperforms the state of the art in terms of predictive performance.

Friederike Baier, Sebastian Mair, Samuel G. Fadel

In contrast to fully-supervised models, self-supervised representation learning only needs a fraction of data to be labeled and often achieves the same or even higher downstream performance. The goal is to pre-train deep neural networks on a self-supervised task, making them able to extract meaningful features from raw input data afterwards. Previously, autoencoders and Siamese networks have been successfully employed as feature extractors for tasks such as image classification. However, both have their individual shortcomings and benefits. In this paper, we combine their complementary strengths by proposing a new method called SidAE (Siamese denoising autoencoder). Using an image classification downstream task, we show that our model outperforms two self-supervised baselines across multiple data sets and scenarios. Crucially, this includes conditions in which only a small amount of labeled data is available. Empirically, the Siamese component has more impact, but the denoising autoencoder is nevertheless necessary to improve performance.

Maria Bånkestad, Jennifer R. Andersson, Sebastian Mair et al.

Reducing a graph while preserving its overall properties is an important problem with many applications. Typically, reduction approaches either remove edges (sparsification) or merge nodes (coarsening) in an unsupervised way with no specific downstream task in mind. In this paper, we present an approach for subsampling graph structures using an Ising model defined on either the nodes or edges and learning the external magnetic field of the Ising model using a graph neural network. Our approach is task-specific as it can learn how to reduce a graph for a specific downstream task in an end-to-end fashion without requiring a differentiable loss function for the task. We showcase the versatility of our approach on four distinct applications: image segmentation, explainability for graph classification, 3D shape sparsification, and sparse approximate matrix inverse determination.

Ahcène Boubekki, Samuel G. Fadel, Sebastian Mair

This paper introduces Neuro-Activated Vision Explanations (NAVE), a method for extracting and visualizing the internal representations of vision model encoders. By clustering feature activations, NAVE provides insights into learned semantics without fine-tuning. Using object localization, we show that NAVE's concepts align with image semantics. Through extensive experiments, we analyze the impact of training strategies and architectures on encoder representation capabilities. Additionally, we apply NAVE to study training artifacts in vision transformers and reveal how weak training strategies and spurious correlations degrade model performance. Our findings establish NAVE as a valuable tool for post-hoc model inspection and improving transparency in vision models.

Ahcène Boubekki, Samuel G. Fadel, Sebastian Mair

Saliency methods have become standard in the explanation toolkit of deep neural networks. Recent developments specific to image classifiers have investigated region-based explanations with either new methods or by adapting well-established ones using ad-hoc superpixel algorithms. In this paper, we aim to avoid relying on these segmenters by extracting a segmentation from the activations of a deep neural network image classifier without fine-tuning the network. Our so-called Neuro-Activated Superpixels (NAS) can isolate the regions of interest in the input relevant to the model's prediction, which boosts high-threshold weakly supervised object localization performance. This property enables the semi-supervised semantic evaluation of saliency methods. The aggregation of NAS with existing saliency methods eases their interpretation and reveals the inconsistencies of the widely used area under the relevance curve metric.

Samuel G. Fadel, Sebastian Mair, Ricardo da S. Torres et al.

Generative models based on normalizing flows are very successful in modeling complex data distributions using simpler ones. However, straightforward linear interpolations show unexpected side effects, as interpolation paths lie outside the area where samples are observed. This is caused by the standard choice of Gaussian base distributions and can be seen in the norms of the interpolated samples as they are outside the data manifold. This observation suggests that changing the way of interpolating should generally result in better interpolations, but it is not clear how to do that in an unambiguous way. In this paper, we solve this issue by enforcing a specific manifold and, hence, change the base distribution, to allow for a principled way of interpolation. Specifically, we use the Dirichlet and von Mises-Fisher base distributions on the probability simplex and the hypersphere, respectively. Our experimental results show superior performance in terms of bits per dimension, Fréchet Inception Distance (FID), and Kernel Inception Distance (KID) scores for interpolation, while maintaining the generative performance.