Shrinking Deep Learning Models: Techniques and Energy Considerations

Explore the world of deep learning model optimization in this 41-minute conference talk from the 37th Chaos Communication Congress (37C3). Gain insights into the energy consumption of modern machine learning models and discover innovative techniques for shrinking their size without compromising performance. Learn about the challenges posed by the end of Moore's law and how neural network parameter counts continue to grow exponentially. Examine various methods for managing model complexity, including low-bitwidth integer representation, pruning redundant connections, and knowledge distillation. Investigate the potential of running cutting-edge language models on consumer-grade GPUs and understand the implications for accessibility and experimentation. Acquire the knowledge needed to make informed decisions about the usage and regulation of deep learning models, while considering their environmental impact and resource requirements.