Fault-tolerant implementation of finite-state automata in recurrent neural networks

摘要

Any deterministic finite-state automata (DFA) can be implemented in a sparse recurrent neural network (RNN) with second-order weights and sigmoidal discriminant functions. Construction algorithms can be extended to fault-tolerant DFA implementations such that faults in an analog implementation of neurons or weights do not affect the desired network performance. The weights are replicated k times for k-1 fault tolerance. Alternatively, the independent network is replicated 2k+1 times and the majority of the outputs is used for a k fault tolerance. In a further alternative solution, a single network with k&eegr; neurons uses a "n choose k"encoding algorithm for k fault tolerance.