LettersAnalog on-chip-learning for active noise canceling
Introduction
Many applications of neural networks require real-time adaptation, which may need specific hardware implementations. Although analog neuro-chips have potential advantages over integration density and speed over digital chips [5], they suffer from nonideal characteristics of the fabricated chips such as offsets and nonlinearity [2], and the fabricated chips are usually not flexible enough to be used for many different applications. Therefore, it is very important to find proper applications where analog neuro-chips have potential advantages over popular digital signal processors (DSPs). Such applications are those with analog input/output (I/O) signals and high computational requirements. For example, intelligent sensors [7], adaptive equalizers [1], [3], and active noise controls are good applications for analog neuro-chips. Also, neuro-chips with on-chip learning capability are essential for such practical applications.
In this Letter a demonstration of analog neuro-chips is reported for active noise canceling. Unlike our previous implementations for adaptive equalizers with binary output [1], both input and output values are analog in this noise canceling application, and more accurate computation is required.
Section snippets
Analog on-chip-learning neuro-chips
The analog neuro-chips developed incorporate error back-propagation learning [1]. Each synapse cell has one capacitor for storing an analog synaptic weight w and three 4-quadrant Gilbert multipliers, one for signal feedforward wx, another for error backpropagation δw, and the other for synaptic weight adjustment δx. Here, x and δ denote the input and the backpropagating delta, respectively. Although the floating-gate technology provides nonvolatile memory, simple capacitors are used here for
Active noise canceling system
As shown in Fig. 1, an active noise canceling (ANC) system generates electric signals to a loudspeaker, which creates acoustic signals to cancel the noise in a quiet zone. Noises propagate from a source to the quiet zone with multiple reflections, characteristics of which are modeled as a finite impulse response (FIR) filter with additive random noises. In general the electric-to-acoustic signal transfer characteristics of the loudspeaker are nonlinear, and the multi-layer perceptron has a
Conclusion
In this Letter we report experimental results of active noise canceling using analog neuro-chips with on-chip learning capability. Although its performance is limited due to both nonideal characteristics of analog circuits and the low resolution of the interface board in these experiments, it clearly demonstrates the feasibility of analog chips for real-world applications with analog input and output values.
Acknowledgements
This research was supported by Korean Ministry of Science and Technology.
References (7)
- J.W. Cho, S.Y. Lee, Analog neuro-chips with on-chip learning capability for adaptive nonlinear equalizer, Electron....
- et al.
Effects of multiplier offsets on on-chip learning for analog neurochip
Neural Process. Lett.
(1996) - et al.
A programmable analog VLSI neural network processor for communication receivers
IEEE Trans. Neural Network
(1993)