A5102713154- Domain Adaptation and Few-Shot Learning
- Generative Adversarial Networks and Image Synthesis
- Human Pose and Action Recognition
- Music and Audio Processing
- Metabolomics and Mass Spectrometry Studies
- Speech Recognition and Synthesis
- Advanced Neural Network Applications
- Neural dynamics and brain function
- Neural Networks and Applications
- Adversarial Robustness in Machine Learning
- Image Processing and 3D Reconstruction
- Advanced Electron Microscopy Techniques and Applications
- Environmental DNA in Biodiversity Studies
- Advanced Statistical Modeling Techniques
- Multimodal Machine Learning Applications
- RNA and protein synthesis mechanisms
- Advanced Vision and Imaging
- Speech and Audio Processing
- CRISPR and Genetic Engineering
- COVID-19 diagnosis using AI
- Cultural Heritage Materials Analysis
- AI in cancer detection
- Identification and Quantification in Food
- Machine Learning and ELM
- Species Distribution and Climate Change
Brain (Germany)
2024
Federal Institute for Risk Assessment
2024
École Polytechnique Fédérale de Lausanne
2021-2024
Helmholtz Zentrum München
2024
Airbus (Germany)
2023
University of Tübingen
2008-2021
CURE International UK
2021
Technical University of Munich
1983-2020
Max Planck Institute for Intelligent Systems
2020
Harvard University
2020
We explore unsupervised pre-training for speech recognition by learning representations of raw audio.wav2vec is trained on large amounts unlabeled audio data and the resulting are then used to improve acoustic model training.We pre-train a simple multi-layer convolutional neural network optimized via noise contrastive binary classification task.Our experiments WSJ reduce WER strong character-based log-mel filterbank baseline up 36 % when only few hours transcribed available.Our approach...
Abstract Estimating the pose of multiple animals is a challenging computer vision problem: frequent interactions cause occlusions and complicate association detected keypoints to correct individuals, as well having highly similar looking that interact more closely than in typical multi-human scenarios. To take up this challenge, we build on DeepLabCut, an open-source estimation toolbox, provide high-performance animal assembly tracking—features required for multi-animal Furthermore,...
We propose vq-wav2vec to learn discrete representations of audio segments through a wav2vec-style self-supervised context prediction task. The algorithm uses either gumbel softmax or online k-means clustering quantize the dense representations. Discretization enables direct application algorithms from NLP community which require inputs. Experiments show that BERT pre-training achieves new state art on TIMIT phoneme classification and WSJ speech recognition.
Abstract Mapping behavioural actions to neural activity is a fundamental goal of neuroscience. As our ability record large and data increases, there growing interest in modelling dynamics during adaptive behaviours probe representations 1–3 . In particular, although latent embeddings can reveal underlying correlates behaviour, we lack nonlinear techniques that explicitly flexibly leverage joint behaviour uncover 3–5 Here, fill this gap with new encoding method, CEBRA, jointly uses...
Today's state-of-the-art machine vision models are vulnerable to image corruptions like blurring or compression artefacts, limiting their performance in many real-world applications. We here argue that popular benchmarks measure model robustness against common (like ImageNet-C) underestimate (but not all) application scenarios. The key insight is scenarios, multiple unlabeled examples of the available and can be used for unsupervised online adaptation. Replacing activation statistics...
Neural networks are highly effective tools for pose estimation. However, as in other computer vision tasks, robustness to out-of-domain data remains a challenge, especially small training sets that common real-world applications. Here, we probe the generalization ability with three architecture classes (MobileNetV2s, ResNets, and EfficientNets) We developed dataset of 30 horses allowed both "within-domain" "out-of-domain" (unseen horse) benchmarking-this is crucial test current human...
Quantification of behavior is critical in diverse applications from neuroscience, veterinary medicine to animal conservation. A common key step for behavioral analysis first extracting relevant keypoints on animals, known as pose estimation. However, reliable inference poses currently requires domain knowledge and manual labeling effort build supervised models. We present SuperAnimal, a method develop unified foundation models that can be used over 45 species, without additional labels....
Multicolored gene reporters for light microscopy are indispensable biomedical research, but equivalent genetic tools electron (EM) still rare despite the increasing importance of nanometer resolution reverse engineering molecular machinery and reliable mapping cellular circuits. We here introduce fully encapsulin/cargo system Quasibacillus thermotolerans (Qt), which in combination with recently characterized encapsulin from Myxococcus xanthus (Mx) enables multiplexed reporter imaging via...
The neural activity of the brain is intimately coupled to dynamics body. Yet how our hierarchical sensorimotor system dynamically orchestrates generation movement while adapting incoming sensory information remains unclear. In mice, extent encoding from posture muscle-level features across motor (M1) and primary forelimb (S1) cortex these are shaped during learning unknown. To address this, we built a large-scale model that captures hypothesized computations use this control novel 50-muscle...
Learning generative object models from unlabelled videos is a long standing problem and required for causal scene modeling. We decompose this into three easier subtasks, provide candidate solutions each of them. Inspired by the Common Fate Principle Gestalt Psychology, we first extract (noisy) masks moving objects via unsupervised motion segmentation. Second, are trained on background objects, respectively. Third, foreground combined in conditional "dead leaves" model to sample novel...
The Direct Current Mode Stirred method (DCMS) has already been outlined in several papers before and is described shortly here for consistency. In principle it combines the of Injection (DCI), which injects currents on surfaces test objects Reverberation Chamber (RC) Radio Frequency (RF) susceptibility testing. motivation mainly driven by high efficiency these two methods therefore resulting field levels, can be generated moderate RF power requirements. major drawback RC, with lowest usable...