The digital transformation we are experiencing in recent years is cross-cutting to all sectors of the society. In the industrial scenario, this transformation is leading towards the fourth industrial revolution characterized by i) large amounts of data collected and ii) decentralization of computational resources along the production line. In this context the use of artificial intelligence (AI) is often subordinated to the adoption of distributed solutions characterized by the use of limited hardware capacity. In this paper, we describe a new framework for learning neural networks on devices with limited resources. A first experimentation on MNIST datasets confirms the validity of the approach that allows to effectively reduce the size of the network during training without significant losses of its accuracy.