A system is fault tolerant if errors or faults are in some way managed to avoid catastrophic outcomes. This applies within AI algorithms; for example, systems with distrubuted knowledge represention, such as the weights in a neural network, have a natural level of fault tolerance in that slight variations of input can still be recognised. It also applies at the level of the architectures supporting AI. Notably deep learning and other big data techniques may require long-running computation in many machines. Frameworks such as MapReduce need to deal with the failure of one or more processors without needing to completely restart the computation.
Used in Chap. 6: pages 74, 86; Chap. 8: page 113
Also known as fault tolerance