Continual Learning in Recurrent Neural Networks for the Internet of Things: A Stochastic Approach

In many applications Internet of Things (IoT) supports decision taking on the base of continuous data acquisition. These data, usually streams of sensed data, are processed and analysed to produce high-level information. The latter task is usually achieved by means of artificial intelligence technologies. Among these, continual learning is emerging as a paradigm that combines well with IoT as it matches the ability of IoT to continuously produce new data. In this context, we address continual learning with Recurrent Neural Networks (RNN) under a stochastic perspective, in which we consider the RNN as a stationary state-space network. This led us to deploy the Generalized Expectation-Maximization algorithm, in a setting suitable for IoT. We demonstrate the effectiveness of our approach by considering a case study taken from digital agriculture, in which we adopt the continual learning model to assess the biomass prediction in the field of horticulture using IoT technology. Results demonstrate that RNNs embedded in the EM framework can learn on their own after a very short training phase covering a few time samples.