Special Report: Optical Patterns

When the optical neurons are combined with both recognition and reconstruction parts, which were demonstrated above, in a simple two-lens system, a compact, ultrafast, and highly efficient neural network system may be implemented in the future. Furthermore, the bulk lenses can be replaced by planar zone plates, allowing the integration of the whole system into the smallest scale possible.

12.5.4 Holographic associative memory with time-division multiplexing In previous work on holographic memory readout with a SELDA, it has been shown that the memory is capable of retrieving a 2-D information with approximately 105 bits within 1 ns, with a total storage capacity of 1011 - 1012 bits in 1 cm3. However, the question is how to detect such ultrafast information, 1014 bits/s or 100 Tbits/s. This is certainly difficult, first because the speed of the current 2-D image sensor is much slower than one billion frames pre second, and second because the light emitted from each microlaser for only 1 ns has to be distributed over a 2-D image, creating a light budget problem. However, whether one can detect it or not, the retrieved information exists in the output plane. So the question is how to utilize that fast information. Fig. 12.16 shows one possible example of utilizing the ultrafast information with currently available devices. In this figure, an input transparency is placed on the plane where memories are retrieved, and the light passing through the input is collected by a single detector with a lens to obtain the inner product of the input and the retrieved memory. The speed of the single detector matches that of a microlaser. Furthermore, because a significant portion of the light from a microlaser is focused to the single detector, serious light budget problems can be avoided.

Fig. 12.16. Time division multiplexing-based holographic associative memory with a SELDA.

Such a TDM system can be extended to an associative memory by the addition of a feedback loop, as in Fig. 12.16. Let us assume that a whole encyclopedia is recorded in the volume holographic memory and partial information on one of the pages is presented in the input plane. As each microlaser is turned on, the corresponding page is retrieved and is instantly compared with the input. As soon as the memory that matches the input appears, the inner product value reaches above a certain threshold level and latches the microlaser, The light from the matched microlaser reconstructs the complete information on the corresponding page on the output plane. In this way, the advantages of optics can be fully utilized even with currently available devices.

12.6 Integration and packaging

To answer the issue of how to align and package a complex optical system in free space optics, a planar integration technology has been developed [86-90]. As can be seen in Fig. 12.17, the free-space coherent optical processor in Fig. 12.17 (a) can be integrated to a planar module as shown in Fig. 12.17 (b).

Previous    Next    Table of Content for report: Optical Patterns    Home

Optical Patterns

 

 

 

Photuris.com - Optical Data Networking