Edge AI (Artificial Intelligence) has become important in embedded systems because it enables AI algorithms to run directly on resource constrained devices. However, there are still several challenges to face when deploying neural networks on edge. For instance, while edge devices present strict constraints in terms of memory, energy and computation, state of the art DNNs have millions of parameters and require massive computational volumes. This makes the original model too slow for latency-critical applications and too large for memory-constrained hardware. To bridge this gap, model compression techniques, such as quantization, can be used to reduce memory demand and speed up inference, but often cause a severe degradation of the final accuracy, especially in low bit-width.