Skip to main content

Reconfigurable optical processing for training and execution of large scale neural networks


The goal of the ROPE project is to develop a re-configurable (programmable) large-scale optical computing platform specifically for the training of large-scale feed-forward multi-layer neural networks. Such networks are commonly exploited in systems where adaptive control and behaviour is required but where the system itself is not easily modelled using standard analytic techniques (i.e. where system identification is problematic). Training these networks is often difficult due to the commonly slow rate of convergence of the weights and the sensitivity of the system to the choice of initial conditions and the parameters of the learning rules. This results either in the choice of inappropriately small networks or the extensive use of a high-performance computer to effect the training. Rope proposes an alternative strategy which will exploit reconfigurable optical processing to implement the network during the training phase and which, because of its speed of processing and interconnect communication, will enable large nets to be trained effectively and which will allow the user to vary the initial conditions and learning parameters to get the best from the system.

The approach will be to exploit newly-developed acousto-optic processing elements which enable a signal represented by an acoustic wave to multiply an incoming light signal. These elements can be organised either in a 1-D array or in a 2-D array and, because of their extremely fast processing speeds: as a result they are very suitable for the implementation of multi-layer perceptron-like nets. The acoustic signal, which represents the network weights (and also the neural connectivity pattern), is generated under the control of a conventional digital computer and, thus, allows both network re-configuration and arbitrary selection of learning model (i.e. the weight-modification algorithm).

The key innovation in this LTR project will be development and evaluation of newly-available re-configurable optical processing components. Once the feasibility of the use of acousto-optic processing elements has been proven, the industrial partners intend to take the work forward to complete the RTD and ultimately to develop a commercial system for reconfigurable optical processing for training and execution of large scale neural networks.


National University Ireland, Maynooth

Participants (2)

Isle Optics
United Kingdom
Monk's Diary Isle Brewers
TA3 6QL Taunton
United Kingdom
The University Of Kent At Canterbury
CT2 7NZ Canterbury