Seminar Data Analytics I: Resource-efficient Deep Learning

Deep Convolutional Neural Networks (CNNs) and Visual Transformers (ViTs) achieve excellent results in many areas of computer vision but face challenges in real-world applications, such as dependence on costly annotated training data and the high amount of computational resources required for training and deploying such models. While the development of recognition algorithms was primarily driven by high recognition rates on large and cleanly annotated datasets for a long time, goals such as learning with fewer computational- and data resources rapidly gain importance.

This seminar focuses on neural network architectures, learning methods, and research areas in the field of deep learning for dealing with low-resource conditions and data scarcity with a special focus on resource-efficient visual recognition. Students will engage in reading state-of-the-art papers, comparing them, and discussing the results.

The seminar will cover, among other topics:
- Low-resource Architectures: Exploration of neural network architectures specifically designed to operate under limited computational resources.
- Neural Network Optimization, Pruning, and Quantization: Techniques to reduce the size and computational demand of neural networks without significantly compromising performance.
- Few-shot Visual Recognition: Methods enabling models training with very few examples.
- Zero-shot Learning and Knowledge Transfer from Language Models: Approaches to leveraging pre-trained vision-language models to address new visual recognition tasks with zero training examples.
- Learning from Simulations and Generative Models for Data Generation: Utilization of simulations and generative models to create synthetic data.