Neural Turing Machines (NTM) combine neural networks with an external memory that functions like the tape of a classical Turing machine.
The neural network controls how data is read, written, and organized in the memory, enabling the model to learn complex algorithmic tasks.
Unlike conventional neural networks, NTMs have the ability to store intermediate results and flexibly access them, which is advantageous for sequential or logical problems.
Training is usually done through the differentiability of all operations, allowing learning via backpropagation on how to use the memory efficiently.
Typical applications include sorting or copying tasks, but also more complex processes like executing simple programs.