Even if it were possible for a genetic code to convey the complexity of the end product, which is isn't [Cha96, pp. 9, 15], it would be highly inefficient. Nature solves this by encoding growth patterns, not the end product. Boers and Kuiper [BK92, Chapter 4] emphasize that the recipe is genetically stored, not the blueprint (also [Cha96, pp. 9, 15, 16]). They model this recipe in terms of L-systems which can approximate growth3.8.
It is well known that modularization has occurred in many biological systems3.9. The brain is not an exception. Modularization plays an essential role, for stability and also for efficiency (modules can be repeated) [Aza00]. The brain can be considered modular as well. The importance of a `bootstrapping process' is highlighted by Metta et.al.
The modules are plastic in critical periods and become more stable over time. The development of modules later on is highly affected by earlier modules. [MPMS00b]
``[Newborns] show a series of `innate' behaviors, basic control synergies and reflexes3.10.'' [MPMS00b]. The primitive reflexes are controlled by the lower `subcortical' areas of the brain and are lost3.11 once the higher centers of the cerebral cortex mature and begin to guide voluntary behaviors [Sha02, p. 137].
The robot explores and exploits its environment simultaneously. When eye movements become reliable and consistent the neck started moving as well, which provided feedback from the inertial sensors. The goal of
the designer has shifted to devising a suitable initial state
(at time ), and the appropriate developmental rules to get
some close approximation of the desired `final product' as opposed to building the final product itself [MPMS00b]. The robot faced problems of overshooting targets and corrective oscillations. The time slots for enabling modules were explicitly but carefully programmed [MPMS00a, p. 9].
This research provides valuable insights into the process of learning of (spatial) perception and originating abilities. For further work it would be interesting to know whether the timing of this `bootstrap process' is genetically orchestrated (literally prescribed) or whether the timings purely depends on the reliability and consistency of a developing module (a dynamic and variant property of individuals). The above experiment3.12 does not yet validate whether this is the case, but makes it plausible. Moreover, neuronal growth mechanisms are `challenged' to create a representation (exploration) of a real-world problem domain, while concurrently exploiting it (making use of what is learned). ``As we examine, [development] is a uniquely powerful and general learning strategy that undermines the central assumptions of classical learnability theory, which is premised on the assumption that the learning properties of a system can be deduced from a fixed computational architecture.'' [QS82, p. 23]
Erik de Bruijn