unidirectional RNN does not consider the whole sequence, but bidirectional RNN consider the whole sequence. Using Viterbi, structured learning consider the whole sequence. Cost and error in RNN are not always related. Cost in structured learning is the upper bound of error. RNN can be deep, but structured learning is hard to deep other than linear structured learning.
單向RNN不考慮整個序列,而雙向RNN考慮整個序列。使用Viterbi,結構化學習考慮整個序列。RNN中的成本和錯誤並不總是相關的。結構學習的成本是誤差的上界。RNN可以很深,但是除了線性結構學習之外,結構學習很難深入。