The ever growing realism and quality of generated videos makes it increasingly harder for humans to spot deepfake content, who need to rely more and more on automatic deepfake detectors. However, deepfake detectors are also prone to errors, and their decisions are not explainable, leaving humans vulnerable to deepfake-based fraud and misinformation. To this end, we introduce ExDDV, the first dataset and benchmark for Explainable Deepfake Detection in Video. ExDDV comprises around 5.4K real and deepfake videos that are manually annotated with text descriptions (to explain the artifacts) and clicks (to point out the artifacts). We evaluate a number of vision-language models on ExDDV, performing experiments with various fine-tuning and in-context learning strategies. Our results show that text and click supervision are both required to develop robust explainable models for deepfake videos, which are able to localize and describe the observed artifacts. Our novel dataset and code to reproduce the results are available at https://github.com/vladhondru25/ExDDV.

Accurate fault classification in rotating machinery under changing speeds and loads is a critical challenge in industrial predictive maintenance, where vibration signatures shift across operating regimes and black-box decisions are difficult to trust. This paper presents a hybrid architecture that combines Transformers with Logic Tensor Networks (LTNs), used here as the neuro-symbolic learning framework because they ground first-order rules into differentiable satisfiability terms optimized directly in the training objective, for fault diagnosis on two public benchmarks: the Drivetrain Dynamics Simulator (DDS) (multiple speed/load regimes) and the University of Connecticut (UoC) gear-fault dataset. A compact 1-D Transformer encodes raw vibration windows, and an LTN layer imposes soft first-order constraints during training. We introduce a dynamic rule module that induces, merges, and prunes centroid-based similarity rules as the embedding geometry evolves, enabling the constraint set to adapt to within-class variability. Unlike prior LTN-based approaches such as LogicLSTM, which reweight a fixed rule set, our rules are induced and updated dynamically during training. Experiments show improvements over strong neural and neuro-symbolic baselines on DDS (average accuracy 94.01% vs 88.20%), and gains over a strong Transformer baseline on UoC (macro F1 0.939). Beyond accuracy, the induced rules provide compact, queryable explanations by identifying prototypical vibration-window patterns that support a prediction. Confidence calibration improves versus baselines under the same evaluation protocol. Because LTN supervision acts only during training, inference latency matches the base Transformer. The results support neuro-symbolic fusion as a practical path to accurate and explainable fault diagnosis under varying operating conditions.

Abstract The third most prevalent cancer nowadays is colorectal cancer. Colonoscopy is an important procedure in the stage of detection of polyps’ malignancy because it helps in early identification and establishes effective therapy. This paper explores specific deep-learning architectures for the binary classification of colorectal polyps and considers the evaluation of their premalignancy risk. The main scope is to create a custom-based deep learning architecture that classifies adenomatous, hyperplastic, and serrated polyps’ samples into benign and premalignant based on images from the colonoscopic dataset. Each image’s output is modified through masked autoencoders which enhance the classification performance of the proposed model, called Bionnica . From the four evaluated state-of-the-art deep learning models (ZF NET, VGG-16, AlexNet, and ResNet-50), our experiments showed that ResNet-50 and ZF NET are most accurate (above 84%), with ResNet-50 excelling at indicating patients with premalignant colorectal polyps (above 92%). ZF NET is the fastest at handling 700 images. Our proposed deep learning model, Bionnica , is more performant than ZF NET and provides an efficient classification of colorectal polyps given its simple structure. The advantage of our model comes from the custom enhancement interpretability with a rule-based layer that guides the learning process and supports medical personnel in their decisions.

Incremental learning enables artificial agents to learn from sequential data. While important progress was made by exploiting deep neural networks, incremental learning remains very challenging. This is particularly the case when no memory of past data is allowed and catastrophic forgetting has a strong negative effect. We tackle class-incremental learning without memory by adapting prediction bias correction, a method which makes predictions of past and new classes more comparable. It was proposed when a memory is allowed and cannot be directly used without memory, since samples of past classes are required. We introduce a two-step learning process which allows the transfer of bias correction parameters between reference and target datasets. Bias correction is first optimized offline on reference datasets which have an associated validation memory. The obtained correction parameters are then transferred to target datasets, for which no memory is available. The second contribution is to introduce a finer modeling of bias correction by learning its parameters per incremental state instead of the usual past vs. new class modeling. The proposed dataset knowledge transfer is applicable to any incremental method which works without memory. We test its effectiveness by applying it to four existing methods. Evaluation with four target datasets and different configurations shows consistent improvement, with practically no computational and memory overhead.

The letter presents a non-massive parallel procedure to compute the biorthogonal dual system used for signal reconstruction in the case of spline-type spaces with multiple generators. The basis of this algorithm are the properties of the projection operator and the invertibility of the Gramian in the case of a Riesz basis. We use a parallel approach in both time and frequency for the computation of the dual system obtained by translation and sampling of a finitely number of atoms. Since there are many applications in signal and image processing where the spline-type spaces (also known as shift-invariant spaces) play a central role, fast computing methods are needed, especially in the multi-windows case, where the computations are expensive from the execution time and from memory storage point of view. We test the implementation on car crash data.