Zero-shot learning holds tremendous potential for histopathology image analysis by enabling models to generalize to unseen classes without extensive labeled data. Recent advancements in vision-language models (VLMs) have expanded the capabilities of ZSL, allowing models to perform tasks without task-specific fine-tuning. However, applying VLMs to histopathology presents considerable challenges due to the complexity of histopathological imagery and the nuanced nature of diagnostic tasks. In this paper, we propose a novel framework called Multi-Resolution Prompt-guided Hybrid Embedding (MR-PHE) to address these challenges in zero-shot histopathology image classification. MR-PHE leverages multiresolution patch extraction to mimic the diagnostic workflow of pathologists, capturing both fine-grained cellular details and broader tissue structures critical for accurate diagnosis. We introduce a hybrid embedding strategy that integrates global image embeddings with weighted patch embeddings, effectively combining local and global contextual information. Additionally, we develop a comprehensive prompt generation and selection framework, enriching class descriptions with domain-specific synonyms and clinically relevant features to enhance semantic understanding. A similarity-based patch weighting mechanism assigns attention-like weights to patches based on their relevance to class embeddings, emphasizing diagnostically important regions during classification. Our approach utilizes pretrained VLM, CONCH for ZSL without requiring domain-specific fine-tuning, offering scalability and reducing dependence on large annotated datasets. Experimental results demonstrate that MR-PHE not only significantly improves zero-shot classification performance on histopathology datasets but also often surpasses fully supervised models.
Current practices for reporting the level of differential privacy (DP) guarantees for machine learning (ML) algorithms provide an incomplete and potentially misleading picture of the guarantees and make it difficult to compare privacy levels across different settings. We argue for using Gaussian differential privacy (GDP) as the primary means of communicating DP guarantees in ML, with the full privacy profile as a secondary option in case GDP is too inaccurate. Unlike other widely used alternatives, GDP has only one parameter, which ensures easy comparability of guarantees, and it can accurately capture the full privacy profile of many important ML applications. To support our claims, we investigate the privacy profiles of state-of-the-art DP large-scale image classification, and the TopDown algorithm for the U.S. Decennial Census, observing that GDP fits the profiles remarkably well in all three cases. Although GDP is ideal for reporting the final guarantees, other formalisms (e.g., privacy loss random variables) are needed for accurate privacy accounting. We show that such intermediate representations can be efficiently converted to GDP with minimal loss in tightness.
We propose a general framework called VisionLogic to extract interpretable logic rules from deep vision models, with a focus on image classification tasks. Given any deep vision model that uses a fully connected layer as the output head, VisionLogic transforms neurons in the last layer into predicates and grounds them into vision concepts using causal validation. In this way, VisionLogic can provide local explanations for single images and global explanations for specific classes in the form of logic rules. Compared to existing interpretable visualization tools such as saliency maps, VisionLogic addresses several key challenges, including the lack of causal explanations, overconfidence in visualizations, and ambiguity in interpretation. VisionLogic also facilitates the study of visual concepts encoded by predicates, particularly how they behave under perturbation -- an area that remains underexplored in the field of hidden semantics. Apart from providing better visual explanations and insights into the visual concepts learned by the model, we show that VisionLogic retains most of the neural network's discriminative power in an interpretable and transparent manner. We envision it as a bridge between complex model behavior and human-understandable explanations, providing trustworthy and actionable insights for real-world applications.
Federated learning (FL) enables the collaborative training of deep neural networks across decentralized data archives (i.e., clients) without sharing the local data of the clients. Most of the existing FL methods assume that the data distributed across all clients is associated with the same data modality. However, remote sensing (RS) images present in different clients can be associated with diverse data modalities. The joint use of the multi-modal RS data can significantly enhance classification performance. To effectively exploit decentralized and unshared multi-modal RS data, our paper introduces a novel multi-modal FL framework for RS image classification problems. The proposed framework comprises three modules: 1) multi-modal fusion (MF); 2) feature whitening (FW); and 3) mutual information maximization (MIM). The MF module employs iterative model averaging to facilitate learning without accessing multi-modal training data on clients. The FW module aims to address the limitations of training data heterogeneity by aligning data distributions across clients. The MIM module aims to model mutual information by maximizing the similarity between images from different modalities. For the experimental analyses, we focus our attention on multi-label classification and pixel-based classification tasks in RS. The results obtained using two benchmark archives show the effectiveness of the proposed framework when compared to state-of-the-art algorithms in the literature. The code of the proposed framework will be available at https://git.tu-berlin.de/rsim/multi-modal-FL.
Classifying images with an interpretable decision-making process is a long-standing problem in computer vision. In recent years, Prototypical Part Networks has gained traction as an approach for self-explainable neural networks, due to their ability to mimic human visual reasoning by providing explanations based on prototypical object parts. However, the quality of the explanations generated by these methods leaves room for improvement, as the prototypes usually focus on repetitive and redundant concepts. Leveraging recent advances in prototype learning, we present a framework for part-based interpretable image classification that learns a set of semantically distinctive object parts for each class, and provides diverse and comprehensive explanations. The core of our method is to learn the part-prototypes in a non-parametric fashion, through clustering deep features extracted from foundation vision models that encode robust semantic information. To quantitatively evaluate the quality of explanations provided by ProtoPNets, we introduce Distinctiveness Score and Comprehensiveness Score. Through evaluation on CUB-200-2011, Stanford Cars and Stanford Dogs datasets, we show that our framework compares favourably against existing ProtoPNets while achieving better interpretability. Code is available at: https://github.com/zijizhu/proto-non-param.
Neural architectures tend to fit their data with relatively simple functions. This "simplicity bias" is widely regarded as key to their success. This paper explores the limits of this principle. Building on recent findings that the simplicity bias stems from ReLU activations [96], we introduce a method to meta-learn new activation functions and inductive biases better suited to specific tasks. Findings: We identify multiple tasks where the simplicity bias is inadequate and ReLUs suboptimal. In these cases, we learn new activation functions that perform better by inducing a prior of higher complexity. Interestingly, these cases correspond to domains where neural networks have historically struggled: tabular data, regression tasks, cases of shortcut learning, and algorithmic grokking tasks. In comparison, the simplicity bias induced by ReLUs proves adequate on image tasks where the best learned activations are nearly identical to ReLUs and GeLUs. Implications: Contrary to popular belief, the simplicity bias of ReLU networks is not universally useful. It is near-optimal for image classification, but other inductive biases are sometimes preferable. We showed that activation functions can control these inductive biases, but future tailored architectures might provide further benefits. Advances are still needed to characterize a model's inductive biases beyond "complexity", and their adequacy with the data.
Efficient training of artificial neural networks remains a key challenge in deep learning. Backpropagation (BP), the standard learning algorithm, relies on gradient descent and typically requires numerous iterations for convergence. In this study, we introduce Expectation Reflection (ER), a novel learning approach that updates weights multiplicatively based on the ratio of observed to predicted outputs. Unlike traditional methods, ER maintains consistency without requiring ad hoc loss functions or learning rate hyperparameters. We extend ER to multilayer networks and demonstrate its effectiveness in performing image classification tasks. Notably, ER achieves optimal weight updates in a single iteration. Additionally, we reinterpret ER as a modified form of gradient descent incorporating the inverse mapping of target propagation. These findings suggest that ER provides an efficient and scalable alternative for training neural networks.
Whole-slide image classification represents a key challenge in computational pathology and medicine. Attention-based multiple instance learning (MIL) has emerged as an effective approach for this problem. However, the effect of attention mechanism architecture on model performance is not well-documented for biomedical imagery. In this work, we compare different methods and implementations of MIL, including deep learning variants. We introduce a new method using higher-dimensional feature spaces for deep MIL. We also develop a novel algorithm for whole-slide image classification where extreme machine learning is combined with attention-based MIL to improve sensitivity and reduce training complexity. We apply our algorithms to the problem of detecting circulating rare cells (CRCs), such as erythroblasts, in peripheral blood. Our results indicate that nonlinearities play a key role in the classification, as removing them leads to a sharp decrease in stability in addition to a decrease in average area under the curve (AUC) of over 4%. We also demonstrate a considerable increase in robustness of the model with improvements of over 10% in average AUC when higher-dimensional feature spaces are leveraged. In addition, we show that extreme learning machines can offer clear improvements in terms of training efficiency by reducing the number of trained parameters by a factor of 5 whilst still maintaining the average AUC to within 1.5% of the deep MIL model. Finally, we discuss options of enriching the classical computing framework with quantum algorithms in the future. This work can thus help pave the way towards more accurate and efficient single-cell diagnostics, one of the building blocks of precision medicine.
DINO and DINOv2 are two model families being widely used to learn representations from unlabeled imagery data at large scales. Their learned representations often enable state-of-the-art performance for downstream tasks, such as image classification and segmentation. However, they employ many empirically motivated design choices and their training pipelines are highly complex and unstable -- many hyperparameters need to be carefully tuned to ensure that the representations do not collapse -- which poses considerable difficulty to improving them or adapting them to new domains. In this work, we posit that we can remove most such-motivated idiosyncrasies in the pre-training pipelines, and only need to add an explicit coding rate term in the loss function to avoid collapse of the representations. As a result, we obtain highly simplified variants of the DINO and DINOv2 which we call SimDINO and SimDINOv2, respectively. Remarkably, these simplified models are more robust to different design choices, such as network architecture and hyperparameters, and they learn even higher-quality representations, measured by performance on downstream tasks, offering a Pareto improvement over the corresponding DINO and DINOv2 models. This work highlights the potential of using simplifying design principles to improve the empirical practice of deep learning.
Bag-based Multiple Instance Learning (MIL) approaches have emerged as the mainstream methodology for Whole Slide Image (WSI) classification. However, most existing methods adopt a segmented training strategy, which first extracts features using a pre-trained feature extractor and then aggregates these features through MIL. This segmented training approach leads to insufficient collaborative optimization between the feature extraction network and the MIL network, preventing end-to-end joint optimization and thereby limiting the overall performance of the model. Additionally, conventional methods typically extract features from all patches of fixed size, ignoring the multi-scale observation characteristics of pathologists. This not only results in significant computational resource waste when tumor regions represent a minimal proportion (as in the Camelyon16 dataset) but may also lead the model to suboptimal solutions. To address these limitations, this paper proposes an end-to-end multi-scale WSI classification framework that integrates multi-scale feature extraction with multiple instance learning. Specifically, our approach includes: (1) a semantic feature filtering module to reduce interference from non-lesion areas; (2) a multi-scale feature extraction module to capture pathological information at different levels; and (3) a multi-scale fusion MIL module for global modeling and feature integration. Through an end-to-end training strategy, we simultaneously optimize both the feature extractor and MIL network, ensuring maximum compatibility between them. Experiments were conducted on three cross-center datasets (DigestPath2019, BCNB, and UBC-OCEAN). Results demonstrate that our proposed method outperforms existing state-of-the-art approaches in terms of both accuracy (ACC) and AUC metrics.
When we train models on biased ML datasets, they not only learn these biases but can inflate them at test time - a phenomenon called bias amplification. To measure bias amplification in ML datasets, many co-occurrence-based metrics have been proposed. Co-occurrence-based metrics are effective in measuring bias amplification in simple problems like image classification. However, these metrics are ineffective for complex problems like image captioning as they cannot capture the semantics of a caption. To measure bias amplification in captions, prior work introduced a predictability-based metric called Leakage in Captioning (LIC). While LIC captures the semantics and context of captions, it has limitations. LIC cannot identify the direction in which bias is amplified, poorly estimates dataset bias due to a weak vocabulary substitution strategy, and is highly sensitive to attacker models (a hyperparameter in predictability-based metrics). To overcome these issues, we propose Directional Predictability Amplification in Captioning (DPAC). DPAC measures directional bias amplification in captions, provides a better estimate of dataset bias using an improved substitution strategy, and is less sensitive to attacker models. Our experiments on the COCO captioning dataset show how DPAC is the most reliable metric to measure bias amplification in captions.
Transformers, particularly Vision Transformers (ViTs), have achieved state-of-the-art performance in large-scale image classification. However, they often require large amounts of data and can exhibit biases that limit their robustness and generalizability. This paper introduces ForAug, a novel data augmentation scheme that addresses these challenges and explicitly includes inductive biases, which commonly are part of the neural network architecture, into the training data. ForAug is constructed by using pretrained foundation models to separate and recombine foreground objects with different backgrounds, enabling fine-grained control over image composition during training. It thus increases the data diversity and effective number of training samples. We demonstrate that training on ForNet, the application of ForAug to ImageNet, significantly improves the accuracy of ViTs and other architectures by up to 4.5 percentage points (p.p.) on ImageNet and 7.3 p.p. on downstream tasks. Importantly, ForAug enables novel ways of analyzing model behavior and quantifying biases. Namely, we introduce metrics for background robustness, foreground focus, center bias, and size bias and show that training on ForNet substantially reduces these biases compared to training on ImageNet. In summary, ForAug provides a valuable tool for analyzing and mitigating biases, enabling the development of more robust and reliable computer vision models. Our code and dataset are publicly available at https://github.com/tobna/ForAug.
Test-time adaptation with pre-trained vision-language models, such as CLIP, aims to adapt the model to new, potentially out-of-distribution test data. Existing methods calculate the similarity between visual embedding and learnable class embeddings, which are initialized by text embeddings, for zero-shot image classification. In this work, we first analyze this process based on Bayes theorem, and observe that the core factors influencing the final prediction are the likelihood and the prior. However, existing methods essentially focus on adapting class embeddings to adapt likelihood, but they often ignore the importance of prior. To address this gap, we propose a novel approach, \textbf{B}ayesian \textbf{C}lass \textbf{A}daptation (BCA), which in addition to continuously updating class embeddings to adapt likelihood, also uses the posterior of incoming samples to continuously update the prior for each class embedding. This dual updating mechanism allows the model to better adapt to distribution shifts and achieve higher prediction accuracy. Our method not only surpasses existing approaches in terms of performance metrics but also maintains superior inference rates and memory usage, making it highly efficient and practical for real-world applications.
Vision Transformer models exhibit immense power yet remain opaque to human understanding, posing challenges and risks for practical applications. While prior research has attempted to demystify these models through input attribution and neuron role analysis, there's been a notable gap in considering layer-level information and the holistic path of information flow across layers. In this paper, we investigate the significance of influential neuron paths within vision Transformers, which is a path of neurons from the model input to output that impacts the model inference most significantly. We first propose a joint influence measure to assess the contribution of a set of neurons to the model outcome. And we further provide a layer-progressive neuron locating approach that efficiently selects the most influential neuron at each layer trying to discover the crucial neuron path from input to output within the target model. Our experiments demonstrate the superiority of our method finding the most influential neuron path along which the information flows, over the existing baseline solutions. Additionally, the neuron paths have illustrated that vision Transformers exhibit some specific inner working mechanism for processing the visual information within the same image category. We further analyze the key effects of these neurons on the image classification task, showcasing that the found neuron paths have already preserved the model capability on downstream tasks, which may also shed some lights on real-world applications like model pruning. The project website including implementation code is available at https://foundation-model-research.github.io/NeuronPath/.
Due to their effective performance, Convolutional Neural Network (CNN) and Vision Transformer (ViT) architectures have become the standard for solving computer vision tasks. Such architectures require large data sets and rely on convolution and self-attention operations. In 2021, MLP-Mixer emerged, an architecture that relies only on Multilayer Perceptron (MLP) and achieves extremely competitive results when compared to CNNs and ViTs. Despite its good performance in computer vision tasks, the MLP-Mixer architecture may not be suitable for refined feature extraction in images. Recently, the Kolmogorov-Arnold Network (KAN) was proposed as a promising alternative to MLP models. KANs promise to improve accuracy and interpretability when compared to MLPs. Therefore, the present work aims to design a new mixer-based architecture, called KAN-Mixers, using KANs as main layers and evaluate its performance, in terms of several performance metrics, in the image classification task. As main results obtained, the KAN-Mixers model was superior to the MLP, MLP-Mixer and KAN models in the Fashion-MNIST and CIFAR-10 datasets, with 0.9030 and 0.6980 of average accuracy, respectively.
Despite the potential of federated learning in medical applications, inconsistent imaging quality across institutions-stemming from lower-quality data from a minority of clients-biases federated models toward more common high-quality images. This raises significant fairness concerns. Existing fair federated learning methods have demonstrated some effectiveness in solving this problem by aligning a single 0th- or 1st-order state of convergence (e.g., training loss or sharpness). However, we argue in this work that fairness based on such a single state is still not an adequate surrogate for fairness during testing, as these single metrics fail to fully capture the convergence characteristics, making them suboptimal for guiding fair learning. To address this limitation, we develop a generalized framework. Specifically, we propose assessing convergence using multiple states, defined as sharpness or perturbed loss computed at varying search distances. Building on this comprehensive assessment, we propose promoting fairness for these states across clients to achieve our ultimate fairness objective. This is accomplished through the proposed method, FedISM+. In FedISM+, the search distance evolves over time, progressively focusing on different states. We then incorporate two components in local training and global aggregation to ensure cross-client fairness for each state. This gradually makes convergence equitable for all states, thereby improving fairness during testing. Our empirical evaluations, performed on the well-known RSNA ICH and ISIC 2019 datasets, demonstrate the superiority of FedISM+ over existing state-of-the-art methods for fair federated learning. The code is available at https://github.com/wnn2000/FFL4MIA.
Integrated gradients is prevalent within machine learning to address the black-box problem of neural networks. The explanations given by integrated gradients depend on a choice of base-point. The choice of base-point is not a priori obvious and can lead to drastically different explanations. There is a longstanding hypothesis that data lies on a low dimensional Riemannian manifold. The quality of explanations on a manifold can be measured by the extent to which an explanation for a point lies in its tangent space. In this work, we propose that the base-point should be chosen such that it maximises the tangential alignment of the explanation. We formalise the notion of tangential alignment and provide theoretical conditions under which a base-point choice will provide explanations lying in the tangent space. We demonstrate how to approximate the optimal base-point on several well-known image classification datasets. Furthermore, we compare the optimal base-point choice with common base-points and three gradient explainability models.
Climate change is one of the most pressing challenges of the 21st century, sparking widespread discourse across social media platforms. Activists, policymakers, and researchers seek to understand public sentiment and narratives while access to social media data has become increasingly restricted in the post-API era. In this study, we analyze a dataset of climate change-related tweets from X (formerly Twitter) shared in 2019, containing 730k tweets along with the shared images. Our approach integrates statistical analysis, image classification, object detection, and sentiment analysis to explore visual narratives in climate discourse. Additionally, we introduce a graphical user interface (GUI) to facilitate interactive data exploration. Our findings reveal key themes in climate communication, highlight sentiment divergence between images and text, and underscore the strengths and limitations of foundation models in analyzing social media imagery. By releasing our code and tools, we aim to support future research on the intersection of climate change, social media, and computer vision.
The Mixture-of-Experts (MoE) model has succeeded in deep learning (DL). However, its complex architecture and advantages over dense models in image classification remain unclear. In previous studies, MoE performance has often been affected by noise and outliers in the input space. Some approaches incorporate input clustering for training MoE models, but most clustering algorithms lack access to labeled data, limiting their effectiveness. This paper introduces the Double-stage Feature-level Clustering and Pseudo-labeling-based Mixture of Experts (DFCP-MoE) framework, which consists of input feature extraction, feature-level clustering, and a computationally efficient pseudo-labeling strategy. This approach reduces the impact of noise and outliers while leveraging a small subset of labeled data to label a large portion of unlabeled inputs. We propose a conditional end-to-end joint training method that improves expert specialization by training the MoE model on well-labeled, clustered inputs. Unlike traditional MoE and dense models, the DFCP-MoE framework effectively captures input space diversity, leading to competitive inference results. We validate our approach on three benchmark datasets for multi-class classification tasks.
Deep learning models have an intrinsic privacy issue as they memorize parts of their training data, creating a privacy leakage. Membership Inference Attacks (MIA) exploit it to obtain confidential information about the data used for training, aiming to steal information. They can be repurposed as a measurement of data integrity by inferring whether it was used to train a machine learning model. While state-of-the-art attacks achieve a significant privacy leakage, their requirements are not feasible enough, hindering their role as practical tools to assess the magnitude of the privacy risk. Moreover, the most appropriate evaluation metric of MIA, the True Positive Rate at low False Positive Rate lacks interpretability. We claim that the incorporation of Few-Shot Learning techniques to the MIA field and a proper qualitative and quantitative privacy evaluation measure should deal with these issues. In this context, our proposal is twofold. We propose a Few-Shot learning based MIA, coined as the FeS-MIA model, which eases the evaluation of the privacy breach of a deep learning model by significantly reducing the number of resources required for the purpose. Furthermore, we propose an interpretable quantitative and qualitative measure of privacy, referred to as Log-MIA measure. Jointly, these proposals provide new tools to assess the privacy leakage and to ease the evaluation of the training data integrity of deep learning models, that is, to analyze the privacy breach of a deep learning model. Experiments carried out with MIA over image classification and language modeling tasks and its comparison to the state-of-the-art show that our proposals excel at reporting the privacy leakage of a deep learning model with little extra information.