Advertisement

Conclusions

  • Martin Hänggi
  • George S. Moschytz
Chapter

Abstract

The applicability of CNNs in the areas of image and signal processing only becomes efficient if an appropriate analog hardware implementing the underlying network is available. Beside significant advantages of CNN chips over digital hardware such as processing speed, small size, and power consumption there is one severe drawback, namely their relatively poor accuracy. Robust operation can only be guaranteed for locally regular tasks, which makes local regularity a key concept for any practical applications of CNN chips:
  • • Locally regular tasks can be carried out on virtually all classes of CNN chip prototypes built so far, whether they are implemented as continuous-time or as sampled-data systems, and whether a saturation or a high-gain nonlinearity is used. They are also executable on cellular automata. Locally irregular tasks, however, are inherently sensitive and therefore not guaranteed to run reliably on an analog chip. Consequently, as long as the accuracy of the CNN circuitry is not substantially improved, the dynamics in the linear region cannot be fully exploited. From a practical point of view, it is sensible to design chips with the type of nonlinearity that is easiest to implement.

Keywords

Cellular Automaton Settling Time Cellular Neural Network Template Design Analog Hardware 
These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.

Copyright information

© Springer Science+Business Media Dordrecht 2000

Authors and Affiliations

  • Martin Hänggi
    • 1
  • George S. Moschytz
  1. 1.Swiss Federal Institute of TechnologyZurichSwitzerland

Personalised recommendations