Processors and memories in computers have trade-offs among performance, power, cost (price) and reliability. Conventional computing does not allow errors because of the sequential processing of executing programs. On the other hand, statistical machine learning applications like image recognition and speech recognition tolerate some errors and inaccuracies. Naturally, human recognition is not perfect. We are working on Approximate computing for machine learning that tolerates errors and inaccuracies at the LSI level such as processor and memory. In the application level, on the other hand, machine learning performs recognition accurately in the same way as conventional computing. Toward future AI applications, we are researching Domain-specific computing optimized for each application.