Different methods of learning are used to shape the many layers of the brain. This is not the way that standard applications of artificial intelligence seemingly modeled after the brain work, however. In almost all cases, the learning algorithms used today exhibit one or two basic processes that are capable of dealing with specific problems, such as image recognition or playing a game.
This is not what we as people typically refer to when we use the word “learning”. The human brain is an immense interconnected complex system made of simpler subsystems, each operating in a way that is far beyond the level of complexity at which they are being currently studied. For this reason, neural-based algorithms up to this point have been built using isolated and simplified concepts from real observations of neural activity. So far, these concepts have proven useful for specific tasks that are typically overcome by the lower levels of mental functioning in humans.
If artificial intelligence is expected to scale in a way that resembles something like human-level or even mammalian-level cognitive adaptation, the existence and interaction of various learning methodologies must be present to allow for different levels of the ‘brain’ to recontextualize knowledge learned at other levels. This allows lower-level concepts to be used in the service of new cognitive functions (i.e. functions that were not responsible for storing the original knowledge).
A great example of this is the way our spatial and temporal knowledge of the world is originally used in building state models, or representations of different configurations of the environment. This knowledge eventually benefits goal-orientation by framing our possible actions in terms of their result as well their utility, which assists in action selection by providing indications of the optimal decision to make at any given time.
Before this occurs however, individual states must be modeled via combinations of individual features, a process that primarily occurs hierarchically. That is, the brain structures involved in this process resemble stacked layers that each receive input from previous layers, combining them into more and more complex patterns. This differs from the structure of state models, which are recurrent in their organization and look more like networks or graphs that hierarchies.
Planning in goal-orientation on the other hand involves tree-like structures that branch outward, reflecting alternative paths that lead to various endpoints or target states. These trees are in fact extracted from state model graphs, and contain particular state transitions that lead from an initial state to an objective. Planning is the process of determining which transitions are to be chosen in order to model a path that leads to the achievement of some goal.
Building Intelligence 