ABSTRACT. In the first part of the lecture, I will explore how we teach AI to generate videos without relying on detailed annotations or object specific labels. By training on collections of similar videos – such as faces or human bodies – the model learns to generalize across an entire category. Building on this idea, we developed a Learnable Game Engine (LGE) that learns from simple monocular videos to keep track of scenes and objects and to re render them from different viewpoints. Much like a real game engine, it captures basic physics and logic, allowing users to control the scene or guide virtual agents through high level language instructions. The second part of the lecture turns to the safety and fairness of generative AI. Most existing approaches look only for predefined types of bias, but real world systems can exhibit unexpected ones. To address this, we introduce OpenBias, a method that uncovers and measures previously unknown biases in text to image models without relying on any preset list. Our experiments show that OpenBias aligns well with established methods and with human judgment, offering a more flexible way to assess fairness in generative systems.

ABSTRACT. The study of the growth rate (or entropy, capacity, information content) of formal languages has drawn the attention of the communities of automata theory, symbolic dynamics and ergodic theory, coding theory, Kolmogorov complexity etc. Its interest to the formal methods community is related with quantitative verification problems, where one needs to evaluate the degree of satisfaction of some property in a model, and the growth rate of the set of behaviors in the model corresponding with the desired property represents an alternative to probabilistic approaches, alternative grounded in deep results from the above-cited domains. We propose here a notion of bandwidth, or information content per time unit, which generalizes growth rate to timed languages, and more particularly to timed automata. We show that timed automata can be partitioned in three classes: (1) meager automata, whose bandwidth is O(1), (2) normal automata, whose bandwidth is Θ(log(1/ε)), and (3) obese, whose bandwidth is Θ(1/ε). Additionally, we provide algorithms for computing the bandwidth for each class of automata. The techniques utilized combine the introduction of a pseudo-distance which defines a pseudo-compact space of timed words, semiring presentations of refinements of the region graph of a timed automaton, spectral radius computation of adjacency matrices of graphs, and the Simon factorization theorem. Based on joint work with Eugène Asarin, Aldric Degorre and Bernardo Jacobo Inclán.

ABSTRACT. Bilevel optimization problems model hierarchical decision-making in which an upper-level decision maker optimizes an objective that depends on the optimal response of a lower-level optimizer. When one or both levels involve multiple conflicting objective functions, the problem becomes a Multi-Objective Bilevel Optimization (MOBO) problem. Evolutionary approaches to this problem class universally adopt Pareto dominance at both levels as the solution criterion — a purely efficiency-based relation that identifies non-dominated solutions without regard for the distribution of objective values across components. In this article we study Lorenz dominance — a refinement of Pareto dominance that additionally favors equitable solutions — as an alternative dominance relation for bilevel multi-objective optimization. Two related research questions are addressed: (i) how Lorenz dominance applied independently and jointly at the upper and lower levels affects the quality of the upper-level front; and (ii) whether the Lorenz-efficient lower-level front constitutes a more realistic basis for the selection process than the full Pareto front, given that classical optimistic and pessimistic selection strategies evaluate lower-level solutions exclusively from the upper-level perspective. Three dominance configurations are evaluated under both selection strategies. Experimental results demonstrate that Lorenz-based configurations match the Pareto baseline under optimistic selection, and consistently produce solutions closer to the theoretical optimum under pessimistic selection.

ABSTRACT. This study addresses the min-max formulation of Bellman’s Lost in a Forest problem for triangular forests: a path is sought that minimizes the escape length in its worst starting position and orientation. It maps the worst-case escape length over the entire space of triangle shapes, parameterized by two angles at a fixed area, with the equilateral triangle, the isosceles triangles, and the right triangles as landmarks. The reliability of such a map rests on the inner maximization, the search for the worst starting case of a fixed path. Restricting this search to a fixed region under-reports the escape length and rewards gimmick paths that exploit the unsearched region. The worst case is therefore evaluated by a dense direct search over the whole triangle, sampling its edges and vertices finely because the worst start lies on the boundary, paired with an evolution strategy that searches for the path, re-testing each candidate against its newly found worst start until none worse remains. The equilateral optimum is a three-segment zig-zag, so the path is fixed to three segments; the reported length is the best within this family: exact at the equilateral, where it matches the global optimum, and an upper bound elsewhere. Across the shape space the equilateral triangle and its near-isosceles neighbours escape with a full three segment zig-zag, while obtuse triangles escape within two. The isosceles triangles, where analytic solutions are known, validate the method: the map reproduces the Besicovitch zig-zag near the equilateral and recovers the equilateral escape length to within 0.03% of its known value 0.981981. Re-running the worst-case search at higher density, including randomly sampled starts, moves the escape length by 0.3% on average, indicating the surface is not an artefact of the sampling density.

ABSTRACT. Gradient-descent bit-flipping (GDBF) is a hard decision decoding algorithm designed for Low-Density Parity-Check (LDPC) codes. While it is attractive for practical implementations, it also suffers from performance limitations due to the local optimum traps that are unavoidable. This work addresses the problem of improving decoding performance using a learning-assisted approach in which a model is employed to guide the decoding process when GDBF updates become ineffective. The proposed method combines conventional bit-flipping and selective data-driven intervention while limiting cost overhead. The Monte-Carlo simulation results show improved error-rate performance compared to the standard GDBF decoder.

ABSTRACT. Recreating a specific guitar timbre from a recording is a complex, non-linear optimization problem traditionally requiring expert manual tuning. While existing deep learning solutions offer high-quality “black-box” clones, they lack interpretable parameters for the end-user. We propose an automated white-box framework that integrates a Large Language Model (LLM) for symbolic music generation with a Genetic Algorithm (GA) for virtual amplifier optimization. Our system uses a 9-parameter non-linear VST signal chain and evaluates similarity through a 28-dimensional acoustic fingerprint (MFCCs, spectral contrast, and spectral rolloff). The proposed framework includes a semantic feedback loop: a rule-based system translates acoustic errors back into natural language, enabling the LLM to refine the musical source across iterations. We conducted a comparison study across six models (360M-70B parameters), identifying Llama3 8B as the optimal threshold for structural musical consistency and computing efficiency. Real-world evaluations on a curated corpus of 30 iconic guitar recordings, spanning Thrash Metal, Classic Rock, Grunge, Progressive Metal, Metalcore and Extreme Metal, demonstrate consistent tonal convergence and the independent discovery of musically meaningful characteristics, such as the thrash metal ‘mid-scoop’, providing a transferable and explainable alternative to neural cloning.

ABSTRACT. AI-generated music detectors are commonly evaluated against original songs, but real-world uploads are often remixed, re-encoded, pitch-shifted, or otherwise edited. These edited versions form a difficult negative class: they are not generated by AI, yet they may introduce spectral artifacts that resemble synthetic audio fingerprints. We study this problem as a hard-negative robustness setting for AI-generated music detection, focusing on AI-generated and edited variants derived from the same anchor songs. We compile a YouTube-based dataset of AI, edited, and original variants, using the original tracks only as references, and train a binary AI-versus-edited detector. Audio is processed as 10-second clips and passed as raw waveforms to a pretrained PaSST spectrogram transformer. To reduce leakage, all splits are performed by anchor song. On the held-out test set, the final video-level system achieves 0.811 balanced accuracy. At clip level, AI-generated clips reach an F1-score of 0.836, while edited clips reach a lower F1-score of 0.720. The results suggest that AI-generated music retains detectable fingerprint-like spectral cues beyond ordinary editing, but the lower edited-class F1 score shows that these cues can still overlap with artifacts from edited audio. Grad-CAM visualizations are used to inspect whether high-confidence predictions rely on localized time–frequency regions.

ABSTRACT. Recent advances in generative modeling have significantly increased the realism of synthetic facial images, making automatic deepfake detection an important task. This paper proposes a hierarchical and explainable framework for deepfake face image analysis. In the first stage, a binary classifier distinguishes authentic images from manipulated ones. In the second stage, images predicted as fake are further assigned to one of three manipulation subtypes: diffusion-based synthesis, identity face swapping using SimSwap, and fully synthetic face generation using StyleGAN2. The framework is implemented using convolutional neural network (CNN) backbones initialized from ImageNet-pretrained weights and is evaluated on a custom dataset containing authentic and manipulated facial images. To improve interpretability, the proposed pipeline is complemented with Grad-CAM++ and SHAP-based visual explanations, allowing the analysis of the image regions that contribute most strongly to model decisions. The proposed approach is intended to support both authenticity detection and manipulation-source attribution within a unified experimental setting.

ABSTRACT. Speech Emotion Recognition (SER) systems built on handcrafted acoustic features, such as Mel-Frequency Cepstral Coefficients (MFCCs), are known to struggle when evaluated out- side their training domain, the cross-corpus generalization gap. We benchmark four self-supervised/weakly-supervised speech models, Wav2Vec2, HuBERT, WavLM, and Whisper, against MFCC-based baselines (SVM, MLP) for in-corpus and cross- corpus SER on RAVDESS, CREMA-D, and EmoDB, evaluating all six pairwise cross-corpus directions for every model. Hu- BERT achieves the best in-corpus Weighted Accuracy (WA) on RAVDESS (99.07%) and CREMA-D (78.96%); WavLM leads on EmoDB (98.77%). All four SSL models outperform the best classical baseline by 10–38 percentage points on every corpus, and degrade far less under cross-corpus transfer (e.g. HuBERT re- tains 63.54% WA transferred CREMA-D→RAVDESS vs. 18.37% for the SVM). We further identify a class-specific generaliza- tion failure: models fine-tuned on the small, German-language EmoDB corpus recognize the “sad” class almost perfectly in- corpus (recall ≥0.98) but collapse to 5–13% sad recall under cross-corpus transfer to RAVDESS and CREMA-D, with sad utterances mostly redirected to “happy”. This collapse is specific to EmoDB as the source corpus: sad recall stays high (0.87–1.00) in every direction where the larger RAVDESS or CREMA-D corpora are the source instead, including transfer into EmoDB, which points to EmoDB’s small size and low-arousal acted portrayals rather than an intrinsic acoustic property of sadness. Layer-wise probing shows emotion-relevant information peaking in the upper transformer layers for all four backbones on all three corpora, with Whisper’s shallower 6-layer encoder saturating earlier and at a consistently lower ceiling than the three 12-layer, raw-waveform models.

ABSTRACT. This study examines how inner-GA hyperparameters selected by a Meta-GA vary with objective, seeded starting condition, mutation operator, evaluation budget, and outer population size. Experiments cover Rosenbrock 2D and 10D, Ackley 2D, and Easom 2D, with three starting points, three evaluation budgets, two mutation operators, and 30 inner repetitions per evolved configuration. Each result is compared with the best of 30 configurations sampled uniformly from the same discrete search space. Meta-GA obtained lower mean fitness in all 216 matched comparisons, with 191 remaining significant after Holm correction. Starting condition most clearly structured the Rosenbrock rankings, whereas evaluation budget most clearly structured the Ackley and Easom rankings. Near-valley Rosenbrock conditions generally selected small mutation scales and small populations with many generations, while harder starts and lower budgets were associated with larger scales. Descriptive analysis of the outer-population-100 configurations found the greatest gene stability for Rosenbrock 10D. The findings support condition-specific tuning but do not establish causal parameter effects or a validated configuration heuristic.

ABSTRACT. Autism Spectrum Disorder (ASD) often impairs the development of daily living skills, requiring structured, repeatable training to foster personal autonomy. This paper presents a virtual reality (VR) application deployed on the Meta Quest 3 that simulates a grocery shopping activity to train daily living skills in children with ASD. The system offers two difficulty levels, an adaptive assistance mechanism, and multimodal visual–auditory feedback to support a gradual transition from guided to more independent task execution.

To enable objective progress monitoring, the VR client continuously records session telemetry, including completion time, correct selections, errors, undo actions, and help requests. These data are transmitted via REST endpoints to a FastAPI backend and stored in a relational database, where they are aggregated into a quantitative performance score that balances execution speed and accuracy, inspired by established speed–accuracy and throughput models. A dedicated web dashboard built with Chart.js provides longitudinal visualizations and automatically generated PDF-ready progress reports for therapists and parents.

Functional and integration testing confirmed reliable synchronization between the VR client, backend services, and analytics dashboard. Overall, the platform shows high feasibility and strong potential to effectively support the training of daily living skills as a complementary therapeutic tool for ASD.

ABSTRACT. The communication of scientific research to a broader audience is difficult because academic papers are usually written for specialists, while producing accessible video content requires additional time and technical expertise. This paper describes a web-based system that automatically converts scientific articles into news-style video presentations. The process starts from a PDF article together with multimedia files provided by the user. The system generates a dialog-based script, produces speech using personalized voices, animates virtual presenters, selects and displays relevant figures from the article, and assembles these components into the final video. Subtitles, background music, and user-selected visual templates can also be included during video generation. The platform also creates a thumbnail, title, description, and tags that can be used to publish the generated video on YouTube. The proposed system combines components for the generation of conversation, speech, and video within a single workflow, with the goal of reducing the manual work required to create audiovisual presentations of scientific papers. The system was tested using scientific articles with different content and visual elements. The results show that the entire video generation workflow can be automated and suggest that the proposed approach can support the communication of scientific research through accessible audiovisual content.

ABSTRACT. This paper investigates model composition—often referred to as "gluing"—within the framework of matching logic. Specifically, we examine the systematic combination of existing signatures, variable valuations, theories, and their corresponding models. Our primary objective is to ensure that this composition preserves satisfaction proofs; that is, any theory validated by the individual constituent models is also validated by the resulting composite model. Our definitions are based on a polyadic, sorted variant of matching logic, which has also been expressed in the Rocq proof assistant. Therefore, we outline our definitions with dependent types for this work to serve as a direct blueprint for the mechanization in the short-term future.

ABSTRACT. We present work in progress towards an intuitionistic version of Applicative Matching Logic. We introduce a semantics based on complete Heyting algebras, and propose a proof system which we prove to be sound relative to this semantics.

ABSTRACT. This paper introduces a novel approach to bias handling in EEG based machine learning models by generating synthetic data using Variational Autoencoders (VAEs) enhanced with attention mechanisms. The article focuses on EEG signals, which often suffer from data scarcity, noise, and subject variability, leading to poor generalization in machine learning models. To overcome these drawbacks, we propose X-BrainVAE, a model that generates synthetic EEG data to complement real datasets, improving classification accuracy and reducing bias. The architecture of X-BrainVAE integrates Transformer blocks and Squeeze and Excitation Blocks to recalibrate feature maps, enhancing the model's ability to extract relevant patterns from the data. By using a probabilistic approach to model the latent space, the VAE allows for diverse and realistic data generation, which is essential for training robust EEG classifiers. The experiments conducted demonstrate the effectiveness of X-BrainVAE in generating high quality synthetic data that fills gaps in real datasets. The generated data not only resembles real EEG signals but also significantly improves classification performance across several tasks, particularly in motor imagery and emotion recognition scenarios. When trained on synthetic data, the models exhibited a 71.7% increase in classification accuracy for previously unseen users, indicating improved generalization capabilities.

ABSTRACT. WordPress plugins often implement their own roles, permissions, customer identities, and staff-account models on top of the native WordPress authorisation system. While these plugin-local models enable domain-specific workflows, they can also create unsafe authorisation boundaries in which plugin-level accounts can influence platform-level identities. This paper presents CVE-2026-6741, an improper privilege management vulnerability in the LatePoint appointment-booking plugin for WordPress, affecting versions up to and including 5.4.1. The flaw enabled an authenticated agent-level user to link a LatePoint customer record to an arbitrary WordPress user account, including an administrator account. Under the vulnerable configuration, the customer password-reset workflow could propagate a new password to the associated WordPress user account, escalating privileges to administrator.
We present this vulnerability as an end-to-end case study of AI-assisted source-code auditing. Specialised audit agents reviewed the plugin codebase and constructed a source-to-sink hypothesis involving customer-to-user account linking and password propagation. We did not treat the generated finding as conclusive; instead, we validated it through manual source review, role and capability analysis, controlled proof-of-concept testing, and examination of the security boundary between LatePoint customer records and WordPress user accounts. This work provides a validated case study showing how AI-generated source-to-sink hypotheses can expose logic and authorisation flaws in plugin-local account models when model outputs are treated as hypotheses requiring independent verification. The case also motivates a security takeaway for WordPress plugin development: account-linking fields and cross-model identity mappings should be treated as authority-bearing state, while password-propagation routines should be treated as privileged operations, even when exposed through plugin-local roles or workflows.

ABSTRACT. Behavioural classification of affective player states from in-game telemetry is a useful building block for adaptive game systems, but the resulting models are typically opaque ensembles whose internal logic is hard to inspect. We present an experimental study of classifying engaged versus frustrated play states from eight low-cost behavioural features extracted from a 2D Flappy-Bird-style environment. The pilot dataset comprises 320 self-labelled rounds collected from 8 participants (40 rounds each, balanced classes). Five classifiers are evaluated under both stratified 5-fold cross-validation per participant and leave-one-subject-out (LOSO) cross-validation. Two findings warrant attention. First, at the pilot scale of n = 8 participants, paired Wilcoxon signed-rank tests detect no significant difference between personalised and LOSO accuracy for any of the five classifiers tested (p > 0.31), suggesting that a single general model may generalise to unseen players without per-participant training. Second, a depth-2 decision tree distilled from the Random Forest reproduces ≈ 95% of the ensemble’s predictions with only four IF-THEN rules, achieving cross validation accuracy of 0.869 ± 0.046 – comparable to the 100-tree ensemble at 0.872 ± 0.118 – and runs in median 59 µs per prediction, ∼ 100× faster than the forest. The result connects numeric ensemble classification to a symbolic rule representation suitable for direct embedding in a game loop. The complete dataset, system implementation, and analysis pipeline are released as a reproducibility package.

ABSTRACT. We investigate the semantics of a concurrent language that can be used to model the structure and behavior of spiking neural P systems with mute rules. For the language under investigation, we design and relate an operational semantics and a denotational semantics, using continuations and the mathematical methodology of metric semantics. We establish the formal relationship between the two semantic models and show that the denotational semantics is correct w.r.t. the operational one.

ABSTRACT. Nowadays, code generated by Artificial Intelligence (AI) can be found almost anywhere, in coursework, technical interview assignments, and public code repositories. Telling this type of code apart from human-written code is also becoming much harder. The best-performing detectors today are based on transformers, such as CodeBERT, GraphCodeBERT, and UniXcoder, which have between 125 and 133M parameters. These detectors can reach an F1 score of 0.97 to 0.99 on in-domain data, but these numbers do not survive cosmetic transformations. Changes that a developer can make in seconds, such as identifier renaming or file reformatting, drag their F1 score down by 27 to 42 points. We instead chose to adopt a structural approach that reads the code’s syntax and data flow rather than its surface text. TypeAwareGAT is an AI code detection model with 2.7M parameters that is based on a graph attention network (GAT) and takes as input the fused abstract syntax tree and data-flow (AST+DFG) graph extracted from the code. Each node contains only the AST type, three structural scalars, and nine lexical features extracted from the node label. On the CoDet-M4 dataset, our model reaches an F1 score of 0.93 on the raw test split and loses no more than 0.001 of that score under renaming and reformatting, while the transformer-based ones drop to values between 0.57 and 0.72. Despite having roughly 50 times fewer parameters, the structural model’s score is nearly unchanged under these transformations, unlike the transformers.

ABSTRACT. Consumer-grade EEG headsets are increasingly used for brain-computer interfaces (BCIs) outside the shielded environments in which they were first validated, yet how everyday noise degrades their performance is poorly characterised. We ask which everyday noise sources actually break a consumer P300 BCI, and how. In a pre-registered, within-subject study of 41 participants on an eight-channel consumer headset, we compared a controlled baseline against three noise conditions: continuous chewing, interference from a saturated 2.4~GHz Wi-Fi channel, and broadband pink noise. Of the three, only chewing reliably degraded P300 classification; the interference and the acoustic noise were both tolerated. Chewing damaged the system at three stages of the pipeline at once: it destroyed roughly a third of all epochs before classification, nearly doubled the false-positive rate among those that survived, and lowered balanced accuracy on the remainder. The result is a practical noise-robustness profile for consumer BCIs in everyday use, and evidence that the dominant failure mode is mechanical and subject-generated rather than electromagnetic.

ABSTRACT. We propose a Łukasiewicz-Moisil symbolic method for interpretable finite-valued models. Starting from a Łukasiewicz formula, we associate a family of Moisil nuance formulas expressing whether the original value reaches each qualitative threshold of a finite chain. This transforms a many-valued expression into a finite sequence of Boolean components, which can be evaluated in parallel. We first illustrate this construction on a Łukasiewicz formula extracted from a neural network, then we apply the same abstraction to gene regulatory networks. Regulatory strengths are discretized into a finite Łukasiewicz chain, and activation and inhibition are represented by fuzzy operations, building human-readable rules about qualitative gene-expression levels.

ABSTRACT. Breast ultrasound classification is difficult because of low signal-to-noise ratios, speckle, operator variability, and appearance shifts across imaging devices. Building on UltraScanNet, a Mamba-based hybrid backbone for this task, we ask a focused question: can a lighter redesign of the backbone retain its accuracy at reduced cost, and can cheap, training-free language cues add anything on top? We introduce a family of variants, collectively UltraScanNet-PF, that (i) reduce the depth of the heaviest stage (from 11 to 7 blocks), (ii) replace that stage's convolutional-attention and self-attention mixers with state-space (UltraScanUnit) blocks -- retaining self-attention only in one variant for comparison -- (iii) expose multi-scale features through a lightweight pyramid-fusion head, and (iv) optionally apply progressive feature refinement. We evaluate on BUSI (three-class) and BUS-BRA (two-class, a different scanner population) under a controlled protocol with repeated runs, and we ablate multi-scale fusion versus refinement, deep self-attention, initialization (domain-partial versus ImageNet), and data augmentation. At their best configuration the lighter variants match or slightly exceed the heavier UltraScanNet baseline -- the most efficient variant (refinement only) matches U2's top-1 on BUSI and exceeds it on BUS-BRA -- while using 7-19% fewer FLOPs and 9-15% fewer parameters; this variant is also the most accurate of the family. We further find that reintroducing self-attention in the deep stages degrades accuracy, that ImageNet pretraining helps the lighter variants only when strong mixing augmentation is disabled (while consistently harming the attention variant), and that one-shot BI-RADS-style text cues generated by a general-purpose vision -- language model yield no improvement over the visual-only models under a matched comparison. Rather than a new state of the art, our contribution is a careful, reproducible characterization of what helps and what does not for Mamba-based breast ultrasound classification.

ABSTRACT. Searching tens of thousands of radiology images by visual similarity takes around 60 milliseconds per query under brute-force cosine search. We build a cluster tree over Biomed- CLIP embeddings that reduces this to under 2 milliseconds while keeping condition-level precision within 1 percent of the exhaustive baseline. The tree extends the RAPTOR construction with a data-driven fan-out selector, a coherence stopping rule, and an anytime breadth-first schedule. The stopping rule is the most surprising contributor: at threshold 0.75 it discards two-thirds of all nodes (retaining one third, 12,265 of 36,762) and precision goes up, not down, by 0.88 percentage points. On the 87,864-image ROCO corpus, beam search with width 5 achieves between 50 and 123-fold speedup depending on tree configuration. A web visualizer with semantic zoom exposes the cluster structure to clinicians. A series of ablations separates the effect of each design decision.