The pRAM


Logical view of a pRAM

Synaptic Connections

There are 256 pRAMs in a pRAM-256 module.

The output of an individual pRAM is represented by one bit in the serial data stream from pRAM_out. Data from pRAM 0 is indicated by the SOP1 signal. The output of a pRAM is also stored internally in the pRAM-256 so that it can be applied to the inputs of other pRAMs in the module or in neighbouring modules.

pRAM modules can access pRAM data in other pRAM modules by specifying the module number in bits 8 to 10 of an input connection pointer.

Using the connection pointers assigned to each input of each pRAM, the connectivity of a neural network can be specified.

Connection pointers allow any pRAM input to be connected to the output of another pRAM (including itself), an external state input, a logic '1' or a logic '0'.

Reward and penalty inputs are used for training the neural network. These can be connected to other data sources as above. They can also be connected to a global Reward and a global Penalty signal respectively.

Using the connection pointers, a range of neural networks can be built from a pRAM-256 module and a range of training schemes can be implemented.

Making Non-learning pRAMs

Any pRAM within a module may be configured as a non-learning pRAM, irrespective of the state of the Train input. This is achieved by setting the connection pointers for its Reward and Penalty inputs to 'GND'. In this way, no modification of the weights in that pRAM will occur.

Fixed-input pRAMS

One weight within the pRAM may be continuously accessed by setting the connection pointers for its inputs to a value of 'VCC' or 'GND' as desired.

More data

T G Clarkson, August 1995