Light Adaptive Architectures

 

A higher level of adaptation can be incorporated in the architecture and algorithm of a vision chip. This adaptation scheme, which can be called algorithmic adaptation, provides a more systematic way to adaptation than an intuitive one. Algorithmic adaptation should be planned when developing the algorithm [Kobayashi et al. 95b]. In this adaptation mechanism a measure of the outputs of the system is provided, and based on that measure, the biasing conditions are controlled. Depending on the method based on which a new biasing condition is chosen, several mechanisms can be recognized.

• Adaptation based on mean value: This method finds the mean of variables at a level of the hierarchy and feeds back that value to control the bias. Using current mode addition, this method can easily be implemented in VLSI. The function of this method can be formulated as

  

  where are the variables at a level of hierarchy

• Adaptation based on mean error or mean square error: Mean error methods can provide information about the deviation of the variables at a level in the hierarchy. This can also be interpreted as an indication of the level of activity. Therefore, using this method the activity of the system can be controlled. This method is formulated as

  

• Adaptation based on maximum value: This method decides upon the largest value of variables. The hardware realization of this method is known as winner-take-all circuits. The maximum value adaptation is useful to tune the gain of the circuits so that the output of the circuits swing at full scale. This method can be described as

  

Choosing an adaptation mechanism depends on the statistics of the signals at the specific hierarchical level. The function f() is also dependent on the method chosen, and the circuits that are to be controlled.