Multifrequency oscillation learning method for analog neural network: Its implementation in a learning LSI

Multifrequency oscillation learning is a method in which each internal parameter of an analog neural network (NN) is slightly oscillated at a different frequency and the parameters are corrected by directly detecting the effect of these oscillations on the NN output. Compared to the conventional methods, high-speed learning and simplification of learning circuit is possible using this method. We fabricated a learning LSI chip based on the multifrequency oscillation learning method. In preparation for fabricating the LSI, a simulation was performed in order to investigate the effect of the hardware characteristics on learning. A 1.3-μm analog CMOS process was chosen for fabrication of the LSI and it was determined that one LSI could control 20 parameters. By using several LSIs, up to 220 parameters can be controlled. To verify the learning behavior of the new LSI, we performed a learning experiment using it together with an analog NN-LSI. A 3-bit parity check problem learning was completed in 800 μs, which verified the high speed of the device. © 1997 Scripta Technica, Inc. Electron Comm Jpn Pt 3, 80(5): 52–64, 1997