Optical Associatron: An Optical Neural Network
The spatial parallelism of optics is well suited to the massively parallel structure of neural network models. The Optical Associatron is an optical auto-associative recalling system with learning capability where SLMs (Spatial Light Modulators) are used as key devices.
The main feature of this system is realizing learning mechanism in an associative memory model. Orthogonal recollection method is employed for learning mechanism. The representation of positive and negative values is difficult for optical analogue processing where the optical intensity is used as an information carrier. But it is not required for this system to represent negative values and it is also possible to recall patterns which have low orthogonality.
SLMs are used for representing a memory matrix and an input pattern, thus full parallel analogue optical multiplication is established. Electronic circuits take charge of refreshing memory matrix which is difficult to be optically realized.
By the feedback type learning method, non-uniformity of optical devices such as SLMs is adaptively compensated.
The learning capability will be shown in the movies below. The numbers are trial numbers.
- Masatoshi Ishikawa, Naohisa Mukohzaka, Haruyoshi Toyoda, and Yoshiji Suzuki: Optical Associatron - A Simple Model for Optical Associative Memory -, Appl. Opt., Vol.28, No.2, pp.291-301 , 1989