Can BiLSTM layer be used as MDLSTM in Matlab. ?
1 view (last 30 days)
I have to design a deep learning system in which MDLSTM (Multidimention LSTM) is required for text line recognition (word and chacter recognition). The neural network I have designed can recognized the individual characters. Research papers described this algorithm with forward, backward, updown and rightlleft passes of LSTM layers (e.g., infamous Alex Grave papers). Most of them have been implemented in Python. I do not find anything in BiLSTM or LSTM layers to control feedforward or backward passes separately in Matlab (while this is possible in Python). Further many research papers have used term 6-tanh activation function as a separate layers. Any idea if these parameters (4-tanh, cell size, block size, forward and backward passes) can be controlled in BILSTM or LSTM layers. The diagram is from a paper (Graves A.) is attached, where red cirlced terms are difficult for me to define in Matlab. Thanks