A collaborative decentralized machine learning (CDML) architecture can support ML model distributed training and inference across highly heterogeneous and resource-constrained IoT devices, which results in less latency, higher reliability, lower energy consumption, and saving bandwidth resources. With using CDML, spare resources across decentralized IoT devices can be fully used to perform computation-intensive ML tasks collaboratively with high performance. This Recommendation introduces collaborative decentralized machine learning (CDML) for intelligent IoT services, and provides the characteristics and reference architecture of CDML for intelligent IoT services.