Optimal Perception

最近寫了一篇還尚滿意的作業

作業一

1. 請說明一項（除了課堂上所提的之外）多重感官系統比起單一感官系統的優勢，或闡述兩者的差異，並以此設想一個最佳化感官系統的機器人所應具有的特性。

Under the pressure of the second law of thermodynamics, the maintenance of complex systems such as perceptual systems require energy and cost from the organism. Therefore, we would predict that perceptual systems would develop/evolve to maximize the amount of ecologically relevant information gained from the environment for survival while keeping down the cost in doing so. As it is necessary for perceptual systems to provide reliable information corresponding to the actual situation in the real world, we could assume that the more “realistic” the system is in providing information relevant to the organism, the more useful the system would be for survival. It is important that not all information should be represented, as presenting the world as it is in its full reality would introduce irrelevant information, which would be even a greater cost to the organism. Selection of information is thus necessary to lower the cost. However, being realistic, that is, increasing the resolution and accuracy of the features involved, would create cost corresponding to the amount of detailed information obtained. We could therefore, in principle, calculate the optimal balancing point of system reliability where the cost for this certain amount of realistic information is minimally affordable for the organism while the benefits of such reliability is maximized.

Can such reliability be achieved through one single modality only? If there is only one single physical principle from which all features originate, whether it be surface smoothness, light wave absorbency, or flexibility, then capturing this principle and its variability with one modality would capture all the patterns of the world. However, even if all physical phenomena could be reduced to quarks or other basic particles and their physical laws, different living beings have been observed to survive by exploiting different superficial levels of macro-properties instead of directly interacting with such micro particles. One possible explanation is that by controlling and perceiving higher level phenomena such as being able to hold up a rock with one's trunk or seeing a juicy red apple, one need not go through the arduous feat of manipulating or looking at each and every quark to achieve the same affects. Therefore, the fact that different organisms interact with different levels of phenomena shows that multiple perceptual systems are required as there is no longer one single modality that could monitor each level at the same time.

However, there is a difference in having multiple perceptual systems that code for the same thing or each coding a different thing.

In the latter case, having only one or many independent perceptual systems would mean that each balancing point would be independently reached in regard to each system's internal costs and benefits. We can view the information obtained by perceptual systems as “codings” of relevant environmental objects. If each system targets their own unique relevant object in the world, then each object is coded by one system only, and the reliability of such coding would depend on how realistic the one and only system is. However, as an organism has only a constant amount of energy to distribute across systems, some objects would then be more poorly represented than others, and the organism would then be weaker in detecting a crucial part of the environment.

On the other hand, multiple coding organisms target the same relevant object. There are two ways of being redundant. One would be to maximize benefits by combining the maximum strengths of each system, adding them up together to obtain greatest reliability; whereas the other would be to minimize the cost by decreasing the reliability of each involved system until the combined result is minimally useful for survival. The first, unlike the second, doesn't consider the cost involved in maintaining high reliability for each system. Therefore only the second could achieve the balance between cost and benefits. Each system involved could be stronger in representing some aspects of some objects while weaker in others, but as long as the systems could complement each other, the objects could be multiply coded without requiring maximum reliability of each system, but only the maximum reliability combination of the systems as a whole. The total cost of maintaining multiple perceptual systems would be much lower than that of single unique coding organisms. It could even be possible for each system to represent an object poorly, but added together, code the object fairly well enough for survival.

There is another difference between multiple perceptual systems that code via single (independent) or multiple (dependent) coding. The second is an all or none situation where the damage of one system means the end of all representations. True, it might be more advantageous to have low cost multiple coding systems, but it would be less costly if the elimination of one modality would only mean the inability to represent a certain range of objects instead of demolishing all. Therefore, depending on the priority and importance of features, one would predict that different living beings would come up with different combinations of dependent and independent (modular) perceptual systems to balance the risk of damaged code with minimized cost and maximized overall relevant information. There is therefore no “one” single solution to life, but many.

The considerations discussed above would be those of building an optimal machine. One first need to specify which levels of macro-phenomena the machine would be interacting with (specifying the ecological relevant features), and then lay out the perceptual systems that could grasp different aspects of objects in those levels. Then set out the priorities (by means of risk analysis in regard to survival) of objects to be detected, with more independent and costly systems combined with some weaker dependent multiple coding systems to those objects, and only dependent and low cost multiple coding systems to low priority objects.