Neural Turing Machines

Neural Turing Machines extend recurrent network architectures with an external memory bank that can be read from and written to at each time step. This process is fully differentiable and can be trained like any other network. The original paper has some interesting demonstrations of NTMs learning basic algorithms. These notes focus on an explanation of the theory and training setup.

Error! Looks like your browser can't display PDFs. Try using an updated version of Chrome or Firefox.