Photo of the system. There are 1120 VLSI chips which are carried on 56 boards, and there are 1008 neurons and 1020 7-bit synapses for each neuron.

Behavior of each neuron is expressed by a nonlinear differential equation.

These circuits are designed by digital circuits. It was fabricated in 0.7 micron CMOS gate arrays with 250000 gates, about 85% of which could be used.

Time constants can be selected from 64 values ranging from 416 microseconds to 26.2 milliseconds including 0 seconds.

Temporal differentiation can be performed by subtracting output of slow neuron from fast neuron. This is a bandpass filter.

Behavior of winner-take-all network composed of 1007 neurons. In 6 milliseconds, the network was conquered by a winner neuron.

The same winner-take-all network was simulated by a workstation and it took 76 seconds to settle. The PDM system was 10000 times faster than a computer.

Experiment of on-line PCA learning.

A sound is detected by two microphones and the signals are fed to two speakers via respective modifiable weights. The joint distribution of the two signals from the two microphones is shown in the rightmost panel. The orientation of the joint distribution directs to the sound source. When PCA is applied to the two-dimensional signal space, the first principal component corresponds to the sound direction and the second principal component is orithgonal to the first component. The first eigen vector is represented by the weights to Speaker 1 and the second one to Speaker 2.

Photo of FPGA circuits implementing real-time PCA learning network..

Loci of the first (on a unit circle) and the second (degenerated) eigen vectors as the sound direction moved.

Motion of joint distribution while sound direction is moving. (Double click to start.)

Locus of the first principal component while sound is moving. (Double click to start.)

TOP

UP