• DropIn
• Resiliency
• RNN
• Dropout

This thesis presents a novel, principled approach to training recurrent neural networks that are robust to missing part of the input features at prediction time.
By building on the ensembling properties of Dropout regularization, we propose a methodology, named DropIn, which efficiently trains a neural network model as a committee machine of subnetworks, each capable of predicting with a subset of the original input features.
We discuss the application of the DropIn methodology to the most representatives recurrent neural models, ranging from simplest recurrent networks to Reservoir Computing models and targeting applications characterized by input sources that might be unreliable or prone to collect discontinued measurements, leading to missingness in input data (e.g., as in pervasive wireless sensor networks and IoT contexts). We provide experimental assessment using real-world data from ambient assisted living and healthcare application domains, showing how the DropIn methodology allows maintaining predictive performances comparable to those of a model without missing features, even when 20%-50% of the inputs are not available.