<p>Neuromorphic electronics can provide in-memory computing systems with low power consumption by emulating key principles of the brain. However, their practical capabilities are limited by a number of challenges, including device non-ideality, limited training accuracy and insufficient adaptability. Here we explore the development of training approaches for neuromorphic electronics, including digital, mixed-signal and emerging neuromorphic electronics. We examine the characteristics and advantages of different training strategies, including off-chip training with on-chip inference, on-chip training and inference, and hybrid offline–online training strategies. We consider the challenges that must be addressed in terms of advanced training, standardized benchmarks and hardware–software co-design, and highlight applications where such efficient training of neuromorphic electronics could be of particular value.</p>