We describe a laser diode-based optoelectronic implementation of artificial neural networks which utilizes real-time holography in photorefractive crystals. The use of a laser diode light source reduces the system size and power requirements. The holographic material is rhodium- doped BaTiO3 which has enhanced sensitivity at the laser-diode wavelength of 830 nm. A balanced coherent-detection method is used to represent bipolar optical neurons and weights. In addition, by distributing each neuron weight among a set of spatially and angularly distributed gratings using beam fanning, Bragg degeneracy and its associated inter-neuron optical crosstalk is virtually eliminated. The structure of the neural network is programmable and we have implemented a variety of neural networks including backpropagation and Kohonen-style self-organizing maps with up to 10,000 neurons and performance of up to 108 weights processed per second during learning and readout. We also discuss weight decay in photorefractive materials, specifically its relative effects in the neural network and data storage domains.
[1]
M. Klein,et al.
Spectroscopic and photorefractive properties of infrared-sensitive rhodium-doped barium titanate.
,
1994,
Optics letters.
[2]
B H Soffer,et al.
Optical interconnection method for neural networks using self-pumped phase-conjugate mirrors.
,
1991,
Optics letters.
[3]
S. H. Lee,et al.
Incremental recording for photorefractive hologram multiplexing.
,
1991,
Optics letters.
[4]
Y Owechko,et al.
Cascaded-grating holography for artificial neural networks.
,
1993,
Applied optics.
[5]
D. Psaltis,et al.
Holography in artificial neural networks
,
1990,
Nature.