The Research Of Selection Of Controller Of DNCs In QA Field

The ultimate goal of artificial intelligence(AI)is to make a kind of artificial general intelligence(AGI).That is to achieve all needed 'intelligence" with only one algorithm or one model.So far,AI has made amazing progress,especially in such fields:image recognition,voice recognition,machine translation and so on.2016-2017,AlphaGo(with its evolutionary versions:AlphaGo Zero and AlphaZero)from Google broad the world"s horizon,and the public got to know the AI has develop that high.However,AlphaZero could only show its muscles in the field of chess game though its breathtaking power.If you ask 'how about the weather,Today?"AlphaZero is also incapable of action.What matters is the most important abilities of AI--reasoning and memorizing are still challenging tasks in laboratory.Fortunately,technical giants like Google and Facebook have make their endeavors on the direction.Recurrent neural networks(RNN)has been proven Turing completeness in theory.Nevertheless,a series of challenges like gradient vanishing and explosion limit the abilities of RNN.The subsequent version of RNN Long-short memory networks(LSTM)solve the problems at a certain extent.But LSTM could still do nothing when the input sequences increase up to a certain threshold.One way to strengthen RNNs is to expand their memorizing ability.2014,Facebook and Google propose their solutions separately.Google"s neural turing machine(NTM)and facebook"s memory network have taken a solid step forward AGI.Specially,the NTM with its subsequent version differentiable neural computers(DNCs)arise a huge research waves of memory augmented AI.In this paper,we focus on application of DNCs in question and answering(QA)field.Along with bAbi dataset,we dig into the abilities of popular variants of RNNs---LSTM,GRU,SCRN,Delta-RNN---to act as a controller of DNCs,what"s more,apply ratio statistical model as a new index of evaluation,and find LSTM(Long-Short Temporal Memory Networks)is the most proper RNNs for DNC when dealing with QA.I derive the consequence from the aspect of time consuming,accuracy,quality of parameters and the decay rate of loss function and so on.