As the use of large embedding models in recommendation systems and language applications increases, concerns over user data privacy have also risen. DP-SGD, a training algorithm that combines differential privacy with stochastic gradient descent, has been the workhorse in protecting user privacy without compromising model accuracy by much. However, applying DP-SGDnaively to embedding models can destroy gradient sparsity, leading to reduced training efficiency. To address this issue, we present two new algorithms, DP-FEST and DP-AdaFEST, that preserve gradient sparsity during private training of large embedding models. Our algorithms achieve substantial reductions (106) in gradient size, while maintaining comparable levels of accuracy, on benchmark real-world datasets.

We demonstrate the effectiveness of our algorithms on four benchmark datasets for online advertising and natural language understanding.

Faithfulness and interpretability are essential for deploying deep neural networks (DNNs) in safety-critical domains such as medical imaging. B-cos networks offer a promising solution by replacing standard linear layers with a weight-input alignment mechanism, producing inherently interpretable, class-specific explanations without post-hoc methods. While maintaining diagnostic performance competitive with state-of-the-art DNNs, standard B-cos models suffer from severe aliasing artifacts in their explanation maps, making them unsuitable for clinical use where clarity is essential. In this work, we address these limitations by introducing anti-aliasing strategies using FLCPooling (FLC) and BlurPool (BP) to significantly improve explanation quality. Our experiments on chest X-ray datasets demonstrate that the modified $\text{B-cos}_\text{FLC}$ and $\text{B-cos}_\text{BP}$ preserve strong predictive performance while providing faithful and artifact-free explanations suitable for clinical application in multi-class and multi-label settings. Code available at: GitHub repository (url: https://github.com/mkleinma/B-cos-medical-paper).

Large Vision-Language Models (LVLMs) have achieved remarkable success but continue to struggle with object hallucination (OH), generating outputs inconsistent with visual inputs. While previous work has proposed methods to reduce OH, the visual decision-making mechanisms that lead to hallucinations remain poorly understood. In this paper, we propose VaLSe, a Vision-aware Latent Steering framework that adopts an interpretation-then-mitigation strategy to address OH in LVLMs. By tackling dual challenges of modeling complex vision-language interactions and eliminating spurious activation artifacts, VaLSe can generate visual contribution maps that trace how specific visual inputs influence individual output tokens. These maps reveal the model's vision-aware focus regions, which are then used to perform latent space steering, realigning internal representations toward semantically relevant content and reducing hallucinated outputs. Extensive experiments demonstrate that VaLSe is a powerful interpretability tool and an effective method for enhancing model robustness against OH across multiple benchmarks. Furthermore, our analysis uncovers limitations in existing OH evaluation metrics, underscoring the need for more nuanced, interpretable, and visually grounded OH benchmarks in future work. Code is available at: https://github.com/Ziwei-Zheng/VaLSe.

The deployment of deep learning models in critical domains necessitates a balance between high accuracy and interpretability. We introduce WASUP, an inherently interpretable neural network that provides local and global explanations of its decision-making process. We prove that these explanations are faithful by fulfilling established axioms for explanations. Leveraging the concept of case-based reasoning, WASUP extracts class-representative support vectors from training images, ensuring they capture relevant features while suppressing irrelevant ones. Classification decisions are made by calculating and aggregating similarity scores between these support vectors and the input's latent feature vector. We employ B-Cos transformations, which align model weights with inputs to enable faithful mappings of latent features back to the input space, facilitating local explanations in addition to global explanations of case-based reasoning. We evaluate WASUP on three tasks: fine-grained classification on Stanford Dogs, multi-label classification on Pascal VOC, and pathology detection on the RSNA dataset. Results indicate that WASUP not only achieves competitive accuracy compared to state-of-the-art black-box models but also offers insightful explanations verified through theoretical analysis. Our findings underscore WASUP's potential for applications where understanding model decisions is as critical as the decisions themselves.

We introduce a new family of neural network models called Convolutional Dynamic Alignment Networks (CoDA-Nets), which are performant classifiers with a high degree of inherent interpretability. Their core building blocks are Dynamic Alignment Units (DAUs), which linearly transform their input with weight vectors that dynamically align with task-relevant patterns. As a result, CoDA-Nets model the classification prediction through a series of input-dependent linear transformations, allowing for linear decomposition of the output into individual input contributions. Given the alignment of the DAUs, the resulting contribution maps align with discriminative input patterns. These model-inherent decompositions are of high visual quality and outperform existing attribution methods under quantitative metrics. Further, CoDA-Nets constitute performant classifiers, achieving on par results to ResNet and VGG models on e.g. CIFAR-10 and TinyImagenet.

As deep learning has achieved state-of-the-art performance for many tasks of EEG-based BCI, many efforts have been made in recent years trying to understand what have been learned by the models. This is commonly done by generating a heatmap indicating to which extent each pixel of the input contributes to the final classification for a trained model. Despite the wide use, it is not yet understood to which extent the obtained interpretation results can be trusted and how accurate they can reflect the model decisions. In order to fill this research gap, we conduct a study to evaluate different deep interpretation techniques quantitatively on EEG datasets. The results reveal the importance of selecting a proper interpretation technique as the initial step. In addition, we also find that the quality of the interpretation results is inconsistent for individual samples despite when a method with an overall good performance is used. Many factors, including model structure and dataset types, could potentially affect the quality of the interpretation results. Based on the observations, we propose a set of procedures that allow the interpretation results to be presented in an understandable and trusted way. We illustrate the usefulness of our method for EEG-based BCI with instances selected from different scenarios.

Most models of visual attention are aimed at predicting either top-down or bottom-up control, as studied using different visual search and free-viewing tasks. We propose Human Attention Transformer (HAT), a single model predicting both forms of attention control. HAT is the new state-of-the-art (SOTA) in predicting the scanpath of fixations made during target-present and target-absent search, and matches or exceeds SOTA in the prediction of taskless free-viewing fixation scanpaths. HAT achieves this new SOTA by using a novel transformer-based architecture and a simplified foveated retina that collectively create a spatio-temporal awareness akin to the dynamic visual working memory of humans. Unlike previous methods that rely on a coarse grid of fixation cells and experience information loss due to fixation discretization, HAT features a dense-prediction architecture and outputs a dense heatmap for each fixation, thus avoiding discretizing fixations. HAT sets a new standard in computational attention, which emphasizes both effectiveness and generality. HAT's demonstrated scope and applicability will likely inspire the development of new attention models that can better predict human behavior in various attention-demanding scenarios.

Transformers increasingly dominate the machine learning landscape across many tasks and domains, which increases the importance for understanding their outputs. While their attention modules provide partial insight into their inner workings, the attention scores have been shown to be insufficient for explaining the models as a whole. To address this, we propose B-cos transformers, which inherently provide holistic explanations for their decisions. Specifically, we formulate each model component--such as the multi-layer perceptrons, attention layers, and the tokenisation module--to be dynamic linear, which allows us to faithfully summarise the entire transformer via a single linear transform. We apply our proposed design to Vision Transformers (ViTs) and show that the resulting models, dubbed Bcos-ViTs, are highly interpretable and perform competitively to baseline ViTs on ImageNet. Code will be made available soon. However, recently they are often surpassed by transformers (Vaswani et al., 2017), which-- if the current development is any indication-- will replace CNNs for ever more tasks and domains. Transformers are thus bound to impact many aspects of our lives: from healthcare, over judicial decisions, to autonomous driving. Given the sensitive nature of such areas, it is of utmost importance to ensure that we can explain the underlying models, which still remains a challenge for transformers. To explain transformers, prior work often focused on the models' attention layers (Jain & Wallace, 2019; Serrano & Smith, 2019; Abnar & Zuidema, 2020; Barkan et al., 2021), as they inherently compute their output in an interpretable manner. For a detailed discussion, see supplement. These model components are given by: a tokenisation module, a mechanism for providing positional information to the model, multi-layer perceptrons (MLPs), as well as normalisation and attention layers, see Figure 1a. By addressing the interpretability of each component individually, we obtain transformers that inherently explain their decisions, see, for example Figure 1 and Figure 1b. In detail, our approach is based on the idea of designing each component to be dynamic linear, such that it computes an input-dependent linear transform. Böhle et al. (2021; 2022), s.t. it can be summarised by a single linear transform for each input.

We introduce a new family of neural network models called Convolutional Dynamic Alignment Networks (CoDA Nets), which are performant classifiers with a high degree of inherent interpretability. Their core building blocks are Dynamic Alignment Units (DAUs), which are optimised to transform their inputs with dynamically computed weight vectors that align with task-relevant patterns. As a result, CoDA Nets model the classification prediction through a series of input-dependent linear transformations, allowing for linear decomposition of the output into individual input contributions. Given the alignment of the DAUs, the resulting contribution maps align with discriminative input patterns. These model-inherent decompositions are of high visual quality and outperform existing attribution methods under quantitative metrics. Further, CoDA Nets constitute performant classifiers, achieving on par results to ResNet and VGG models on e.g. CIFAR-10 and TinyImagenet. Lastly, CoDA Nets can be combined with conventional neural network models to yield powerful classifiers that more easily scale to complex datasets such as Imagenet whilst exhibiting an increased interpretable depth, i.e., the output can be explained well in terms of contributions from intermediate layers within the network.