I believe that training methods for recursive neural networks will be some of the most interesting future research in the field. We've seen some pretty revolutionary techniques come out for feedforward models in the last 30 years, but with recurrent/recursive networks it's often difficult to see what's going on inside the network, let alone properly train it. We've already seen what Deep RNNs can do when trained [1], but I think this is only the beginning.
Comments
I believe that training methods for recursive neural networks will be some of the most interesting future research in the field. We've seen some pretty revolutionary techniques come out for feedforward models in the last 30 years, but with recurrent/recursive networks it's often difficult to see what's going on inside the network, let alone properly train it. We've already seen what Deep RNNs can do when trained [1], but I think this is only the beginning.
[1] http://arxiv.org/pdf/1303.5778.pdf