Learning and Memory Part #1: Neural Networks and Artificial Intelligence

Since the first speculations about the nature of the human psyche, it has been generally recognized that there were various elements comprising that psyche. Those early approximations have yielded the more explicit concepts of emotion, rationality, and volition, but these categories have proven to be much too broad to accurately account for the range of human experience. In the early part of the twentieth century, psychologists discounted all but the volitive actions that could be perceptibly studied in an attempt to develop a more scientific rigor in a field that has deep roots in philosophy.

There were two major developments that helped to reorient psychology on the internal processes. One was the creation of technological tools for observing the brain in action such as fMRIs and PET scans. The other was the development of computers, and was a major force in the creation of the field of cognitive psychology. It also initiated the search for artificial intelligence, and the question of what exactly that intelligence must entail. One of the key ideas that have come out of both of these fields is the concept of neural networks.

The process of communication within the brain occurs mainly through electrical impulses sent through neural cells. There are an estimated 100 billion nerve cells in the human body, and they are the underlying communication network that allows action, contemplation, and all of the subjective experiences of consciousness. The process by which a nerve cell acts on another nerve cell is incredibly complex, but the essential process involves the opening of ion-channels which create an electrical charge known as an action potential that will travel down the cell's axon towards the dendritic tentacles of another neuron. Once this action potential reaches the end of the axon, it will stimulate synaptic vesicles full of chemicals known as neurotransmitters to release those chemicals across the synaptic gap between the axon and the dendrites of the next nerve cell. These neurotransmitters will either excite or inhibit the opening of ion-channels in the neighboring nerve cell leading to a firing of another action potential if a certain threshold of excitatory neurotransmitter messages is achieved.

There is also an important secondary process that occurs during these reactions in many of these cells that is believed to be the underlying process of memory and learning. While neurotransmitters affect the opening of these ion-channels, they also can set off the release of a second messenger. These messengers set off very complex processes that can help to grow more neuroreceptors and can close the synaptic gap. This strengthening of neurons is implicated in the process of long term potentiation, which allows the flow of communication to be achieved more effectively, and possibly creates the storehouse of memory in these changes at the neurological level.

Clearly, the process of communication between even two nerve cells is incredibly complex, yet these microcosmic processes underlie the macrocosm of the human mind. It is generally accepted that the brain is full of bundles of neurons that act as modules enacting or processing some specific aspect of human experience. Generally these processes have been broken down into four categories. The two overriding categories are cognitive and affective, but both of these can be broken down into controlled and automatic processes. Controlled processes are the serial experiences of consciousness, while the automatic processes are not accessible to consciousness, but are none-the-less crucial to our experiences of consciousness. Within each of these categories there are many modules of neural bundles that perform specialized tasks, and there is both cooperation and competition between these modules for the attention of consciousness and the creation of action.

Both in the early tradition of psychology and in many commonsense expressions of human cognition, people have held that rationality should be the overriding control process, but neuroscience tells us that this is a fallacious view. There are far more neural connections from the limbic system, which is linked to emotionality, to the cerebral cortex, which is linked to thought, so the idea that rationality can control emotion would seem to be troublesome. There is also evidence that the most neural modules feed into both the prefrontal cortex, which is associated with long term planning, and the striatum, which is a part of the basal ganglia and is associated with motor activity. It would seem from this evidence that there is a need for a balancing of the self and a less explicitly rational approach to controlling behavior.

Regardless, it is clear that human cognition and behavior are the result of complex processes of neural modules that can either compete or cooperate in creating the subjective experience of human consciousness. This is a notion that is at the center of a long-standing controversy in the field of Artificial Intelligence, whether or not a machine will ever have subjective experiences. John Searle, a philosopher from MIT, proposed the notion that computers will simply have syntactical knowledge, but never semantic knowledge. He used an argument in which you are in a room with a basket full of blocks with Chinese symbols on them. You are given explicit rules for which symbols to pass out of a window based on the symbols passed in. You may seem to be responding to questions or statements in Chinese, but you wouldn't be said to understand Chinese. Searle claims that this is the limitation of artificial intelligence, and so far his argument stands.

The early field of artificial intelligence focused on creating programs that could learn, known as expert systems. There was some success in this field, especially in the field of game play, such as computer chess players, but these early successes didn't yield much beyond game simulations. Another sub-field in AI was neural network theory, which at first was considered an inferior field because you can't explain how a neural network learns. Neuroscientists also found these neural simulations to be much too simplistic to replicate human action, but they have been successful in things like facial recognition in ways that traditional AI have not.

While both of these fields had there time in the sun, it quickly became clear that a combination of the two would be the most effective at simulating human behavior, and as a result the field of new AI developed. In this field specific modules of codelets, or small pieces of code, can be activated by specific actions by the user. Essentially, this new approach uses psychological constructs, such as working memory, perception, emotion, as the basis for these modules that will then compete for the attention of the "consciousness" of the program for executable behaviors.
These new AI programs can simulate human behavior more realistically, but they are still very situation dependent, in that they can only operate within the confines of a specific function, and rely on our still very fragmented understanding of human psychology. The larger question of whether or not these machines could ever experience some form of consciousness remains to be seen. It would seem that Searle's argument will hold without some unexpected technological shifts, such as the one's that brought about the question in the first place.