Learning effective neural networks on resource-constrained devices

The digital transformation we are experiencing inrecent years is cross-cutting to all sectors of the society. In theindustrial scenario, this transformation is leading towards thefourth industrial revolution characterized by i) large amounts ofdata collected and ii) decentralization of computational resourcesalong the production line. In this context the use of artificialintelligence (AI) is often subordinated to the adoption of distributedsolutions characterized by the use of limited capacityhardware. In this paper we describe a new framework forlearning neural networks on devices with limited resources. Afirst experimentation on MNIST datasets confirms the validityof the approach that allows to effectively reduce the size of thenetwork during training without significant losses of its accuracy.