Modular Concepts

Modeling cognitive processes in a more human-like way is a central concern in artificial intelligence research. The transferability of acquired knowledge across different tasks can be considered a key property of natural intelligence. However, most deep learning approaches implement monolithic systems mixing knowledge and skills. We present an architecture for multi-agent language emergence where we represent knowledge and skills as separate small neural modules. We show that such an architecture can allow a population of agents to bootstrap an efficient shared language grounded in perception and can form a basis for the transferability of knowledge to other tasks.