## Structure

Computer has a CPU and a RAM.

Differential Neural Computer has a neural network as the controller that take the role of the CPU.
The memory is an $N W$ matrix that take the role of *the RAM, where $N$ means the locations and $W$ means the length of each pieces of memory.

## Memory augmentation and attention mechanism

The episodic memories or evenet memories are known to depend on the hippocampus in the human brain.

The main point is that the memory of the network is external to the network itself.

The attention mechanism defines some distributions over the $N$ locations.
Each $i-th$ component of a weighting vector will communicate how much attention the controller should give to the content in the $i-th$ location of the memory.

## Differntiability

Every unit and operation in this structure is differentiable.

## Weightings

The controller wants to do something which involves memory, and it doesn’t just look at every location of the memor.
Instead, it focues its attention on those locations which contain the information it is looking for.

The weighting produced for an input is a distribution over the N locations for their relative importance in a particular process(reading or writing).

Note that the weightings are produced by means by a vector emitted by the controller, which is called interface vector. The

## Three interactions between controller and memory

The controller and memory are mediated by the interface vector.

### Content lookup

A particular set of values within the interface vector, which we will collect in something called key vector, is compared to the content of each location. This comparison is made by means of a similarity measure.

The transitions between consecutively written locations are recorded in an $N * N$ matrix, called temproal link matrix “L”. The sequence by which the controller writes in the memory is an information by itself, and it is something we want to store.