Minimizing inconsistencies across successive versions of an AI system is as crucial as reducing the overall error. In image classification, such inconsistencies manifest as negative flips, where an updated model misclassifies test samples that were previously classified correctly. This issue becomes increasingly pronounced as the number of training classes grows over time, since adding new categories reduces the margin of each class and may introduce conflicting patterns that undermine their learning process, thereby degrading performance on the original subset. To mitigate negative flips, we propose a novel approach that preserves the margins of the original model while learning an improved one. Our method encourages a larger relative margin between the previously learned and newly introduced classes by introducing an explicit margin-calibration term on the logits. However, overly constraining the logit margin for the new classes can significantly degrade their accuracy compared to a new independently trained model. To address this, we integrate a double-source focal distillation loss with the previous model and a new independently trained model, learning an appropriate decision margin from both old and new data, even under a logit margin calibration. Extensive experiments on image classification benchmarks demonstrate that our approach consistently reduces the negative flip rate with high overall accuracy.

Existing deepfake detectors achieve near-perfect accuracy when trained and tested on the same generation method, often without requiring complicated detection pipelines. However, these detectors struggle to generalize well to unseen fake images, because the artifacts on which they rely to distinguish real from fake content are not consistently present or distinctive across varying data distributions. In this study, we focus on understanding where and why detectors fail to generalize in cross-dataset scenarios, leveraging Explainable AI (XAI) methods to identify the specific failure points via a thoroughly feature-level inspection. Based on this analysis, we propose a straightforward recovery method designed to restore detection ability without significantly compromising overall performance; such an approach can be easily integrated into existing detection pipelines as a plug-and-play solution. We use a simple CNN-based synthetic image detector in our experiments for making an understanding of generalization issues and exploring the recovery process. Our approach addresses the generalization gap observed between in-dataset and cross-dataset contexts. Experiments conducted on various generative methods and implementations demonstrate the effectiveness of the proposed recovery strategy.

Small, fast, and lightweight drones present significant challenges for traditional RGB cameras due to their limitations in capturing fast-moving objects, especially under challenging lighting conditions. Event cameras offer an ideal solution, providing high temporal definition and dynamic range, yet existing benchmarks often lack fine temporal resolution or drone-specific motion patterns, hindering progress in these areas. This paper introduces the Florence RGB-Event Drone dataset (FRED), a novel multimodal dataset specifically designed for drone detection, tracking, and trajectory forecasting, combining RGB video and event streams. FRED features more than 7 hours of densely annotated drone trajectories, using 5 different drone models and including challenging scenarios such as rain and adverse lighting conditions. We provide detailed evaluation protocols and standard metrics for each task, facilitating reproducible benchmarking. The authors hope FRED will advance research in high-speed drone perception and multimodal spatiotemporal understanding.

Learning compatible representations enables the interchangeable use of semantic features as models are updated over time. This is particularly relevant in search and retrieval systems where it is crucial to avoid reprocessing of the gallery images with the updated model. While recent research has shown promising empirical evidence, there is still a lack of comprehensive theoretical understanding about learning compatible representations. In this paper, we demonstrate that the stationary representations learned by the d-Simplex fixed classifier optimally approximate compatibility representation according to the two inequality constraints of its formal definition. This not only establishes a solid foundation for future works in this line of research but also presents implications that can be exploited in practical learning scenarios. An exemplary application is the nowstandard practice of downloading and fine-tuning new pretrained models. Specifically, we show the strengths and critical issues of stationary representations in the case in which a model undergoing sequential fine-tuning is asynchronously replaced by downloading a better-performing model pretrained elsewhere. Such a representation enables seamless delivery of retrieval service (i.e., no reprocessing of gallery images) and offers improved performance without operational disruptions during model replacement. Code available at: https://github.com/miccunifi/iamcl2r.

In today's multimedia-rich environment, the rapid growth of data poses significant challenges for developing efficient multi-modal retrieval systems essential for retrieving text, images, audio, and video. As data expands, newer, scalable, and high-performance retrieval systems are increasingly necessary. Embedding-based deep neural networks (DNNs) have become key solutions, transforming high-dimensional data into lower-dimensional embeddings for easy comparison and retrieval. However, updating DNNs changes the internal feature representations, necessitating the extraction of new feature vectors for all gallery data, which is costly, especially with gallery sets comprising billions of data. Learning backward-compatible representations addresses this by allowing new representation to be matched with old gallery data without recalculating features. This tutorial aims to equip participants with the knowledge and tools to apply backward-compatible representations, enhancing multimedia retrieval systems' efficiency and scalability. Participants will learn the importance of compatible representations, basic methods and techniques, and explore challenging open questions that are becoming increasingly relevant to multimedia and cross-modal retrieval.

Compatible features enable the direct comparison of old and new learned features allowing to use them interchangeably over time. In visual search systems, this eliminates the need to extract new features from the gallery-set when the representation model is upgraded with novel data. This has a big value in real applications as re-indexing the gallery-set can be computationally expensive when the gallery-set is large, or even infeasible due to privacy or other concerns of the application. In this paper, we propose CoReS, a new training procedure to learn representations that are compatible with those previously learned, grounding on the stationarity of the features as provided by fixed classifiers based on polytopes. With this solution, classes are maximally separated in the representation space and maintain their spatial configuration stationary as new classes are added, so that there is no need to learn any mappings between representations nor to impose pairwise training with the previously learned model. We demonstrate that our training procedure largely outperforms the current state of the art and is particularly effective in the case of multiple upgrades of the training-set, which is the typical case in real applications.

In this article, we propose a method to partially mimic natural intelligence for the problem of lifelong learning representations that are compatible. We take the perspective of a learning agent that is interested in recognizing object instances in an open dynamic universe in a way in which any update to its internal feature representation does not render the features in the gallery unusable for visual search. We refer to this learning problem as Compatible Lifelong Learning Representations (CL 2 R), as it considers compatible representation learning within the lifelong learning paradigm. We identify stationarity as the property that the feature representation is required to hold to achieve compatibility and propose a novel training procedure that encourages local and global stationarity on the learned representation. Due to stationarity, the statistical properties of the learned features do not change over time, making them interoperable with previously learned features. Extensive experiments on standard benchmark datasets show that our CL 2 R training procedure outperforms alternative baselines and state-of-the-art methods. We also provide novel metrics to specifically evaluate compatible representation learning under catastrophic forgetting in various sequential learning tasks. Code is available at https://github.com/NiccoBiondi/CompatibleLifelongRepresentation .