A5002157796- Adversarial Robustness in Machine Learning
- Anomaly Detection Techniques and Applications
- Advanced Neural Network Applications
- Integrated Circuits and Semiconductor Failure Analysis
- Domain Adaptation and Few-Shot Learning
- Machine Learning and Data Classification
- Explainable Artificial Intelligence (XAI)
- Visual Attention and Saliency Detection
- Advancements in Semiconductor Devices and Circuit Design
- 3D Shape Modeling and Analysis
- Advanced Vision and Imaging
- Face and Expression Recognition
- Artificial Intelligence in Healthcare and Education
- Medical Image Segmentation Techniques
- Physical Unclonable Functions (PUFs) and Hardware Security
- Remote-Sensing Image Classification
- Neural Networks and Applications
- Advanced Image and Video Retrieval Techniques
- Advanced Memory and Neural Computing
- Retinal Imaging and Analysis
- Image Retrieval and Classification Techniques
- Gaussian Processes and Bayesian Inference
- Mathematical Biology Tumor Growth
- Text and Document Classification Technologies
- Ultrasonics and Acoustic Wave Propagation
Google (United Kingdom)
2025
Max Planck Institute for Informatics
2018-2024
Max Planck Society
2018-2021
Central Bank of Luxembourg
2021
RWTH Aachen University
2014-2017
Universidade do Estado do Rio de Janeiro
2005
3D shape completion from partial point clouds is a fundamental problem in computer vision and graphics. Recent approaches can be characterized as either data-driven or learning-based. Data-driven rely on model whose parameters are optimized to fit the observations. Learning-based approaches, contrast, avoid expensive optimization step instead directly predict complete incomplete observations using deep neural networks. However, full supervision required which often not available practice. In...
Obtaining deep networks that are robust against adversarial examples and generalize well is an open problem. A recent hypothesis even states both accurate models impossible, i.e., robustness generalization conflicting goals. In effort to clarify the relationship between generalization, we assume underlying, low-dimensional data manifold show that: 1. regular leave manifold; 2. constrained manifold, on-manifold examples, exist; 3. errors, training boosts generalization; 4. not necessarily...
Excellence in a wide variety of medical applications poses considerable challenges for AI, requiring advanced reasoning, access to up-to-date knowledge and understanding complex multimodal data. Gemini models, with strong general capabilities long-context offer exciting possibilities medicine. Building on these core strengths Gemini, we introduce Med-Gemini, family highly capable models that are specialized medicine the ability seamlessly use web search, can be efficiently tailored novel...
We address the problem of 3D shape completion from sparse and noisy point clouds, a fundamental in computer vision robotics. Recent approaches are either data-driven or learning-based: Data-driven rely on model whose parameters optimized to fit observations; Learning-based approaches, contrast, avoid expensive optimization step by learning directly predict complete shapes incomplete observations fully-supervised setting. However, full supervision is often not available practice. In this...
Despite the success of vision transformers (ViTs), they still suffer from significant drops in accuracy presence common corruptions, such as noise or blur. Interestingly, we observe that attention mechanism ViTs tends to rely on few important tokens, a phenomenon call token overfocusing. More critically, these tokens are not robust often leading highly diverging patterns. In this paper, intend alleviate overfocusing issue and make more stable through two general techniques: First, our...
Adversarial training yields robust models against a specific threat model, e.g., $L_\infty$ adversarial examples. Typically robustness does not generalize to previously unseen models, other $L_p$ norms, or larger perturbations. Our confidence-calibrated (CCAT) tackles this problem by biasing the model towards low confidence predictions on By allowing reject examples with confidence, generalizes beyond employed during training. CCAT, trained only examples, increases $L_\infty$, $L_2$, $L_1$...
Adversarial training (AT) has become the de-facto standard to obtain models robust against adversarial examples. However, AT exhibits severe overfitting: cross-entropy loss on examples, so-called loss, decreases continuously while eventually increasing test In practice, this leads poor generalization, i.e., robustness does not generalize well new paper, we study relationship between generalization and flatness of landscape in weight space, whether changes significantly when perturbing...
Despite their success, vision transformers still remain vulnerable to image corruptions, such as noise or blur. Indeed, we find that the vulnerability mainly stems from unstable self-attention mechanism, which is inherently built upon patch-based inputs and often becomes overly sensitive corruptions across patches. For example, when only occlude a small number of patches with random (e.g., 10%), these patch would lead severe accuracy drops greatly distract intermediate attention layers. To...
<title>Abstract</title> AI systems based on Large Language Models (LLMs) have demonstrated great potential for conducting diagnostic conversations but evaluation has been largely limited to language-only interactions, deviating from the real-world requirements of remote care delivery. Instant messaging platforms can permit clinicians and patients upload discuss multimodal medical artifacts seamlessly in conversation, ability LLMs reason over such data while preserving other attributes...
Deep neural network (DNN) accelerators received considerable attention in recent years due to the potential save energy compared mainstream hardware. Low-voltage operation of DNN allows further reduce consumption, however, causes bit-level failures memory storing quantized weights. Furthermore, are vulnerable adversarial attacks on voltage controllers or individual bits. In this paper, we show that a combination robust fixed-point quantization, weight clipping, as well random bit error...
We develop a principled procedure for determining when large language model (LLM) should abstain from responding (e.g., by saying "I don't know") in general domain, instead of resorting to possibly "hallucinating" non-sensical or incorrect answer. Building on earlier approaches that use self-consistency as more reliable measure confidence, we propose using the LLM itself self-evaluate similarity between each its sampled responses given query. then further leverage conformal prediction...