OPTICS IN 1989

OPTICAL COMPUTING

Optical learning machine for multi-category classification

E.G. Paek, J.R. Wullert II, and J.S. Patel, Bellcore

The recent increased activity in the area of neural net­works is mainly due to their potential to mimic the

computational ability of humans, in particular the ability to learn. While the ordinary computer requires precise rules to perform a given task, the neural computer has the ability to learn its own rules from examples presented to it. This gives the neural computer the potential to solve problems not possible by artificial rule-based systems.

Optics and VLSI are considered to be the most promis­ing technologies for implementing these learning algo­rithms. However, optics has some strong advantages over VLSI, especially for two-dimensional problems, due to its inherent parallelism in two dimensions.

Most of the proposed 1 - 5 optical learning machines are based on the classical holographic set-up with a photore­fractive crystal as a medium to form variable interconnec­tions between inputs and outputs. However, in spite of the number of proposed systems and the current enthusiastic interest in implementing learning machines, the surprising­ly few systems3 demonstrated so far can only classify in­puts into two categories. A multi-category classifier with many output neurons would be an important advance be­cause it could classify inputs into many classes, making it possible to solve a greater range of problems than a two-category classifier only capable of making 'yes/no' deci­sions.

We have demonstrated an optical learning machine that can function as a multi-category classifier. The system is capable of classifying inputs into 2 1 0 = 1024 possible out­put states. We achieve performance by using bipolar inter­connections and outputs, which allows us to implement the perceptron7 learning algorithm, which in turn allows the solution of problems not possible by unipolar repre­sentation.

The system was tested to solve a small scale real-world problem of character recognition. Each of the input train­ing patterns was loaded onto a two-dimensional spatial light modulator and focused onto a photorefractive crys­tal. The light diffracted from the crystal was collected and focused onto a detector array, consisting of 20 elements to represent the positive and negative values for ten outputs. The outputs from each pair of detectors were compared

Schematic of the holographic learning machine for mul­ticategory classification. SLM (spatial light modulator); LCM (liquid crystal modulator); OMDR (optical memo­ry disc recorder).

and the difference was thresholded as required for the per­ception algorithm. The error signals were generated by comparing the output with the desired target signals and were loaded onto a 20-element liquid crystal modulator (LCM), again to represent 10 bipolar error values.

Five standard capital letters were used as training pat­terns. Even though the training patterns have many seg­ments in common, the system successfully converged to the no-error state where all the characters were properly identified. The ultimate capacity of this system has not yet been investigated. Performance can be even more im­proved by cascading.

This is the first multi-category learning machine that can be used for digit recognition.

REFERENCES

1. K. Wagner and D. Psaltis, Appl. Opt. 26, p. 5061, 1987. 2. D.Z. Anderson, Proc. of the ICNN, San Diego, 1987, III, p. 577. 3. D. Psaltis, D. Brady, and K. Wagner, Appl. Opt. 9, p. 1752, 1988. 4. Y. Owechko and B.H. Soffer, Proc. of IEEE International Conference

on Neural Networks, San Diego, Calif., 1988, II, p. 385. 5. J. Hong and Pochi Yeh, Proc. of the Topical Meeting on Optical Com­

puting, Salt Lake, 1989, p. 307. 6. E.G. Paek, J.R. Wullert II, and J.S. Patel, to be published in Opt. Lett.,

Dec. 1, 1989. 7. F. Rosenblatt, Cornell Aeronaut Lab. Report, pp. 85-460, 1, 1987.

28 OPTICS NEWS • DECEMBER 1989