Memory-augmented neural networks

The goal of MANNs is to excel at one-shot learning tasks. The NMT controller, as we read earlier, uses both content-based addressing and location-based addressing. On the other hand, the MANN controller uses only content-based addressing. There are two reasons for this. One reason is that location-based addressing is not required for one-shot learning tasks. In this task, for a given input, there are only two actions that a controller might need to take and both actions are content dependent and not location dependent. One action is taken when the input is very similar to a previously seen input, in which case we can update the current contents of the memory. The other action is taken when the current input is not similar to previously seen inputs, in which case we don't want to overwrite the recent information; instead, we write to the least...