From the Lab to the Edge: Post-Training Compression for Deep Neural Networks

Watch a 58-minute technical talk exploring how Datakalab tackles the challenge of deploying deep neural networks (DNNs) efficiently on edge devices. Learn about a two-step approach that enables framework-agnostic inference support across diverse hardware platforms and implements advanced compression techniques. Discover how post-training quantization, pruning, and context adaptation methods achieve significant model optimization while maintaining accuracy within 1% of the original performance. Presented by PhD student Edouard Yvinec from Sorbonne Université, gain insights into practical solutions for transitioning DNNs from development frameworks like TensorFlow and PyTorch to resource-constrained edge devices without requiring intensive cloud computing or model retraining.