An Internet of Agents Architecture for Training and Deployment of Deep Convolutional Models

It is a fact that Artificial Intelligence is having an ever-growing impact on society. That is not just because of advances in computational power and in machine learning models, such as deep neural networks, but also because of the availability of a large volume of heterogeneous data from diverse sources. The Internet of Things (IoT) paradigm is helping gather massive amounts of data from sensor networks that can be used to train and generate complex AI models. However, the training of these models needs not only the data but has high computational requirements. In this scenario there has appeared a new paradigm, called the Internet of Agents (IoA), which allows the inclusion of intelligence and autonomy in IoT devices and networks. This paper presents an IoA architecture that allows the continual and distributed generation and exploitation of convolutional neural networks. Specific protocols for the safe and efficient transmission of models and training pictures are designed. The convolutional model is trained in the cloud and, once reduced, it is distributed and executed in agents located in embedded devices with low computational resources. The architecture has been tested using a convolutional model for the recognition of handwritten character digits based on the MNIST database.