A5000291599- Domain Adaptation and Few-Shot Learning
- Machine Learning and Data Classification
- Advanced Image and Video Retrieval Techniques
- Generative Adversarial Networks and Image Synthesis
- Advanced Neural Network Applications
- Gaussian Processes and Bayesian Inference
- Adversarial Robustness in Machine Learning
- Human Pose and Action Recognition
- Image Retrieval and Classification Techniques
- Robotics and Sensor-Based Localization
- Machine Learning and Algorithms
- Neural Networks and Applications
- Advanced Vision and Imaging
- Anomaly Detection Techniques and Applications
- Model Reduction and Neural Networks
- Machine Learning in Healthcare
- Medical Image Segmentation Techniques
- Bayesian Modeling and Causal Inference
- Explainable Artificial Intelligence (XAI)
- Image and Signal Denoising Methods
- Multimodal Machine Learning Applications
- Advanced Optical Sensing Technologies
- Machine Learning in Materials Science
- Advanced Image Processing Techniques
- Computer Graphics and Visualization Techniques
Microsoft Research (United Kingdom)
2013-2024
Microsoft (United States)
2010-2021
Google (United States)
2018-2020
Microsoft Research (India)
2016-2017
Max Planck Society
2006-2016
Max Planck Institute for Intelligent Systems
2011-2016
Max Planck Institute for Biological Cybernetics
2006-2011
Shanghai Jiao Tong University
2005-2006
With the advent of deep neural networks, learning-based approaches for 3D reconstruction have gained popularity. However, unlike images, in there is no canonical representation which both computationally and memory efficient yet allows representing high-resolution geometry arbitrary topology. Many state-of-the-art can hence only represent very coarse or are limited to a restricted domain. In this paper, we propose Occupancy Networks, new methods. networks implicitly surface as continuous...
A key ingredient in the design of visual object classification systems is identification relevant class specific aspects while being robust to intra-class variations. While this a necessity order generalize beyond given set training images, it also very difficult problem due high variability appearance within each class. In last years substantial performance gains on challenging benchmark datasets have been reported literature. This progress can be attributed two developments: highly...
Generative neural samplers are probabilistic models that implement sampling using feedforward networks: they take a random input vector and produce sample from probability distribution defined by the network weights. These expressive allow efficient computation of samples derivatives, but cannot be used for computing likelihoods or marginalization. The generative-adversarial training method allows to train such through use an auxiliary discriminative network. We show approach is special case...
Recent work has shown local convergence of GAN training for absolutely continuous data and generator distributions. In this paper, we show that the requirement absolute continuity is necessary: describe a simple yet prototypical counterexample showing in more realistic case distributions are not continuous, unregularized always convergent. Furthermore, discuss regularization strategies were recently proposed to stabilize training. Our analysis shows with instance noise or zero-centered...
RANSAC is an important algorithm in robust optimization and a central building block for many computer vision applications. In recent years, traditionally hand-crafted pipelines have been replaced by deep learning pipelines, which can be trained end-to-end fashion. However, has so far not used as part of such because its hypothesis selection procedure non-differentiable. this work, we present two different ways to overcome limitation. The most promising approach inspired reinforcement...
Modern machine learning methods including deep have achieved great success in predictive accuracy for supervised tasks, but may still fall short giving useful estimates of their {\em uncertainty}. Quantifying uncertainty is especially critical real-world settings, which often involve input distributions that are shifted from the training distribution due to a variety factors sample bias and non-stationarity. In such well calibrated convey information about when model's output should (or not)...
Adversarial perturbations of normal images are usually imperceptible to humans, but they can seriously confuse state-of-the-art machine learning models. What makes them so special in the eyes image classifiers? In this paper, we show empirically that adversarial examples mainly lie low probability regions training distribution, regardless attack types and targeted Using statistical hypothesis testing, find modern neural density models surprisingly good at detecting perturbations. Based on...
Entertainment and gaming systems such as the Wii XBox Kinect have brought touchless, body-movement based interfaces to masses. Systems like these enable estimation of movements various body parts from raw inertial motion or depth sensor data. However, interface developer is still left with challenging task creating a system that recognizes embodying meaning. The machine learning approach for tackling this problem requires collection data sets contain relevant their associated semantic...
Variational Autoencoders (VAEs) are expressive latent variable models that can be used to learn complex probability distributions from training data. However, the quality of resulting model crucially relies on expressiveness inference model. We introduce Adversarial Bayes (AVB), a technique for with arbitrarily models. achieve this by introducing an auxiliary discriminative network allows rephrase maximum-likelihood-problem as two-player game, hence establishing principled connection between...
Deep generative models based on Generative Adversarial Networks (GANs) have demonstrated impressive sample quality but in order to work they require a careful choice of architecture, parameter initialization, and selection hyper-parameters. This fragility is part due dimensional mismatch or non-overlapping support between the model distribution data distribution, causing their density ratio associated f-divergence be undefined. We overcome this fundamental limitation propose new...
Even years ago, Szeliski et al. published an influential study on energy minimization methods for Markov random fields (MRF). This provided valuable insights in choosing the best optimization technique certain classes of problems. While these remain generally useful today, phenominal success field models means that kinds inference problems we solve have changed significantly. Specifically, today often include higher order interactions, flexible connectivity structures, large label-spaces...
Deep generative models based on Generative Adversarial Networks (GANs) have demonstrated impressive sample quality but in order to work they require a careful choice of architecture, parameter initialization, and selection hyper-parameters. This fragility is part due dimensional mismatch or non-overlapping support between the model distribution data distribution, causing their density ratio associated f -divergence be undefined. We overcome this fundamental limitation propose new...
Recent approaches to action classification in videos have used sparse spatio-temporal words encoding local appearance around interesting movements. Most of these use a histogram representation, discarding the temporal order among features. But this ordering information can contain important about itself e.g. consider sport disciplines hurdle race and long jump, where global motions (running, jumping) is discriminate between two. In work we propose sequential representation which retains...
Dynamic Bayesian networks such as Hidden Markov Models (HMMs) are successfully used probabilistic models for human motion. The use of hidden variables makes them expressive models, but inference is only approximate and requires procedures particle filters or chain Monte Carlo methods. In this work we propose to instead simple that model observed quantities. We retain a highly dynamic by using interactions nonlinear non-parametric. A presentation our approach in terms latent shows logarithmic...