Abstract:Masked Image Modeling (MIM) has emerged as a popular method for Self-Supervised Learning (SSL) of visual representations. However, for high-level perception tasks, MIM-pretrained models offer lower out-of-the-box representation quality than the Joint-Embedding Architectures (JEA) - another prominent SSL paradigm. To understand this performance gap, we analyze the information flow in Vision Transformers (ViT) learned by both approaches. We reveal that whereas JEAs construct their representation on a selected set of relevant image fragments, MIM models aggregate nearly whole image content. Moreover, we demonstrate that MIM-trained ViTs retain valuable information within their patch tokens, which is not effectively captured by the global [cls] token representations. Therefore, selective aggregation of relevant patch tokens, without any fine-tuning, results in consistently higher-quality of MIM representations. To our knowledge, we are the first to highlight the lack of effective representation aggregation as an emergent issue of MIM and propose directions to address it, contributing to future advances in Self-Supervised Learning.
Abstract:Deep Learning (DL) models are often black boxes, making their decision-making processes difficult to interpret. This lack of transparency has driven advancements in eXplainable Artificial Intelligence (XAI), a field dedicated to clarifying the reasoning behind DL model predictions. Among these, attribution-based methods such as LRP and GradCAM are widely used, though they rely on approximations that can be imprecise. To address these limitations, we introduce One Matrix to Explain Neural Networks (OMENN), a novel post-hoc method that represents a neural network as a single, interpretable matrix for each specific input. This matrix is constructed through a series of linear transformations that represent the processing of the input by each successive layer in the neural network. As a result, OMENN provides locally precise, attribution-based explanations of the input across various modern models, including ViTs and CNNs. We present a theoretical analysis of OMENN based on dynamic linearity property and validate its effectiveness with extensive tests on two XAI benchmarks, demonstrating that OMENN is competitive with state-of-the-art methods.
Abstract:Prototypical parts networks, such as ProtoPNet, became popular due to their potential to produce more genuine explanations than post-hoc methods. However, for a long time, this potential has been strictly theoretical, and no systematic studies have existed to support it. That changed recently with the introduction of the FunnyBirds benchmark, which includes metrics for evaluating different aspects of explanations. However, this benchmark employs attribution maps visualization for all explanation techniques except for the ProtoPNet, for which the bounding boxes are used. This choice significantly influences the metric scores and questions the conclusions stated in FunnyBirds publication. In this study, we comprehensively compare metric scores obtained for two types of ProtoPNet visualizations: bounding boxes and similarity maps. Our analysis indicates that employing similarity maps aligns better with the essence of ProtoPNet, as evidenced by different metric scores obtained from FunnyBirds. Therefore, we advocate using similarity maps as a visualization technique for prototypical parts networks in explainability evaluation benchmarks.
Abstract:Collaborative self-supervised learning has recently become feasible in highly distributed environments by dividing the network layers between client devices and a central server. However, state-of-the-art methods, such as MocoSFL, are optimized for network division at the initial layers, which decreases the protection of the client data and increases communication overhead. In this paper, we demonstrate that splitting depth is crucial for maintaining privacy and communication efficiency in distributed training. We also show that MocoSFL suffers from a catastrophic quality deterioration for the minimal communication overhead. As a remedy, we introduce Momentum-Aligned contrastive Split Federated Learning (MonAcoSFL), which aligns online and momentum client models during training procedure. Consequently, we achieve state-of-the-art accuracy while significantly reducing the communication overhead, making MonAcoSFL more practical in real-world scenarios.
Abstract:Prototypical parts networks combine the power of deep learning with the explainability of case-based reasoning to make accurate, interpretable decisions. They follow the this looks like that reasoning, representing each prototypical part with patches from training images. However, a single image patch comprises multiple visual features, such as color, shape, and texture, making it difficult for users to identify which feature is important to the model. To reduce this ambiguity, we introduce the Lucid Prototypical Parts Network (LucidPPN), a novel prototypical parts network that separates color prototypes from other visual features. Our method employs two reasoning branches: one for non-color visual features, processing grayscale images, and another focusing solely on color information. This separation allows us to clarify whether the model's decisions are based on color, shape, or texture. Additionally, LucidPPN identifies prototypical parts corresponding to semantic parts of classified objects, making comparisons between data classes more intuitive, e.g., when two bird species might differ primarily in belly color. Our experiments demonstrate that the two branches are complementary and together achieve results comparable to baseline methods. More importantly, LucidPPN generates less ambiguous prototypical parts, enhancing user understanding.
Abstract:Active Visual Exploration (AVE) is a task that involves dynamically selecting observations (glimpses), which is critical to facilitate comprehension and navigation within an environment. While modern AVE methods have demonstrated impressive performance, they are constrained to fixed-scale glimpses from rigid grids. In contrast, existing mobile platforms equipped with optical zoom capabilities can capture glimpses of arbitrary positions and scales. To address this gap between software and hardware capabilities, we introduce AdaGlimpse. It uses Soft Actor-Critic, a reinforcement learning algorithm tailored for exploration tasks, to select glimpses of arbitrary position and scale. This approach enables our model to rapidly establish a general awareness of the environment before zooming in for detailed analysis. Experimental results demonstrate that AdaGlimpse surpasses previous methods across various visual tasks while maintaining greater applicability in realistic AVE scenarios.
Abstract:Partial Multi-label Learning (PML) is a type of weakly supervised learning where each training instance corresponds to a set of candidate labels, among which only some are true. In this paper, we introduce \our{}, a novel probabilistic approach to this problem that extends the binary cross entropy to the PML setup. In contrast to existing methods, it does not require suboptimal disambiguation and, as such, can be applied to any deep architecture. Furthermore, experiments conducted on artificial and real-world datasets indicate that \our{} outperforms existing approaches, especially for high noise in a candidate set.
Abstract:Class-incremental learning is becoming more popular as it helps models widen their applicability while not forgetting what they already know. A trend in this area is to use a mixture-of-expert technique, where different models work together to solve the task. However, the experts are usually trained all at once using whole task data, which makes them all prone to forgetting and increasing computational burden. To address this limitation, we introduce a novel approach named SEED. SEED selects only one, the most optimal expert for a considered task, and uses data from this task to fine-tune only this expert. For this purpose, each expert represents each class with a Gaussian distribution, and the optimal expert is selected based on the similarity of those distributions. Consequently, SEED increases diversity and heterogeneity within the experts while maintaining the high stability of this ensemble method. The extensive experiments demonstrate that SEED achieves state-of-the-art performance in exemplar-free settings across various scenarios, showing the potential of expert diversification through data in continual learning.
Abstract:Vision Transformers (ViTs) overpass Convolutional Neural Networks in processing incomplete inputs because they do not require the imputation of missing values. Therefore, ViTs are well suited for sequential decision-making, e.g. in the Active Visual Exploration problem. However, they are computationally inefficient because they perform a full forward pass each time a piece of new sequential information arrives. To reduce this computational inefficiency, we introduce the TOken REcycling (TORE) modification for the ViT inference, which can be used with any architecture. TORE divides ViT into two parts, iterator and aggregator. An iterator processes sequential information separately into midway tokens, which are cached. The aggregator processes midway tokens jointly to obtain the prediction. This way, we can reuse the results of computations made by iterator. Except for efficient sequential inference, we propose a complementary training policy, which significantly reduces the computational burden associated with sequential decision-making while achieving state-of-the-art accuracy.
Abstract:Vision transformers have excelled in various computer vision tasks but mostly rely on rigid input sampling using a fixed-size grid of patches. This limits their applicability in real-world problems, such as in the field of robotics and UAVs, where one can utilize higher input elasticity to boost model performance and efficiency. Our paper addresses this limitation by formalizing the concept of input elasticity for vision transformers and introducing an evaluation protocol, including dedicated metrics for measuring input elasticity. Moreover, we propose modifications to the transformer architecture and training regime, which increase its elasticity. Through extensive experimentation, we spotlight opportunities and challenges associated with input sampling strategies.