With the penetration of AI technology into society, it has become necessary to consider using the right AI in the right place. Large-scale AI with extremely high-performance capabilities for cloud service consumes a lot of energy and requires a large amount of data for training. In contrast, edge AI placed at the edge of the Internet aims to adaptively process data at high speed using low-power computing devices, potentially mitigating increased data traffic for cloud AI. In this chapter, we first explain that reservoir computing is one of the promising lightweight machine learning frameworks for temporal pattern recognition. Next, we introduce reservoir computing models, including the echo state network and its variant models, with particular focus on their timescale conversion property for time series data. We demonstrate that extended echo state networks with broadly distributed time constants are able to better approximate multi-timescale nonlinear dynamical systems than conventional models. The demonstration suggests an importance of appropriate selection of machine learning models that match the timescale of biological and human-related data targeted by slow electronics.