Data-Efficient Deep Learning
Data is essential to achieve good generalization of deep learning (DL) models. However, many restrictions in the data collection process often lead to the absence of sufficient amount of training data to enable satisfactory performance. Synthetic data generation with deep generative models, such as Generative Adversarial Networks or Diffusion Models, have great potential for data-efficient deep learning. At Bosch IoC Lab, we aim to investigate methods that allow synthesizing data points that are almost indistinguishable from real data. This can help to vastly reduce the need to collect new data and improve the performance of deep learning models trained on both real and synthetic data. Data-efficient methods for deep learning have huge cost saving potential and can also contribute to making algorithms safer and more robust to deal with safety-critical situations, e.g., by synthesizing rare cases or dangerous situations for data augmentation or validation purporses for which real-world data is scarce.
TEAM
Dr. Anna Khoreva
Research Group Leader
Bosch Center for Artificial Intelligence
Research Focus:
Data-Efficient Deep Learning, with a particular focus on Generative Models, Image and Video Synthesis, Few-Shot Learning, Unsupervised and Weakly Supervised Learning
Nikita Kister
PhD Student
Research Focus:
Generative Models, Virtual Humans, Human-Scene Interaction
Co-supervised with Prof. Gerard Pons-Moll
nikita.kister @uni-tuebingen.de
Massimo Bini
PhD Student
Research Focus:
Generative Models, Compositionality
Co-supervised with Prof. Zeynep Akata
massimo.bini @uni-tuebingen.de
PROJECTS
Realistic Synthesis of Humans in Complex and Crowded Scenes / Nikita Kister
Recent advances in deep generative models have led to an unprecedented level of realism in synthesizing human faces and human avatars, reposing humans or swapping their identities and appearance. However, little attention has been paid on synthesizing full body humans in their realistic environments, surrounded by various objects, interacting with other people and performing different tasks. The aim of this project is to address this limitation and to investigate how to augment real-world scenes with synthetic humans. In particular, we aim to investigate how to synthesize realistic humans in complex and crowded scenes based on their 3D body shapes and poses using deep generative models, 3D modeling and neural rendering techniques.
In cooperation with Prof. Gerard Pons-Moll
Compositional Data Synthesis / Massimo Bini
Human cognition is extremely good at extracting generalizable features from limited examples. However, current machine learning models still struggle to generalize, and they often fail when facing data that are not seen during training. In this regard, the ability to synthesize novel unseen data, would allow models to be more robust, providing additional training data that would be difficult or expensive to gather.
One pivotal element behind human generalization capability is the ability to learn complex representations by combining individual concepts. In this regard, we argue that approaching data generation in a compositional manner could be the key to reach human-level results in an automated way.
In cooperation with Prof. Zeynep Akata
Explainable Machine Learning Tübingen (EML)