Memories play an important role in Machine Learning (ML) implementations; they store all model parameters and computation results and thus, the correctness of memory data determines the dependability of the entire ML system. Unfortunately, memories can suffer from different types of soft errors, such as bit-flip errors caused by radiation-induced single event upset, or stuck-at errors due to approximate storage with scaled supply voltage. The use of Error Correction Codes (ECCs) is a typical solution to protect memories against soft errors by introducing some hardware redundancy. However, for ML implementations that usually have a large size of memory while requiring hardware efficiency, there is little room to employ conventional ECC approaches for protection. Another critical dependability issue that is faced by ML systems and often addressed with ECCs is the noise during wireless transmission in IoT applications. In this case, the use of ECC is still challenging because its required power dissipation may not meet the strict requirements of hardware-constraint platforms for IoT devices. Focusing on these problems, this chapter presents some recent hardware-efficient ECC schemes designed for dependable ML systems; by exploring the inherent error tolerance of ML algorithms, these schemes can guarantee dependable operation with very low (or even none in some cases) hardware redundancy and/or power dissipation.