Category Archives: Cognitive Science
The Dynamical Approach
In the field of Artificial Intelligence, it is assumed that the brain is a type of computer: it can read inputs from the senses, represent the world symbolically, compute plans, and take action. Cognitive processes are represented as a kind of flow chart. But does the brain really function like a computer? The computer model does not take into account the fact that the brain activity is a continuous, time-dependent process. Unlike a computer program which is fed a problem, computes an answer at some future time, then waits for the next problem, the human brain may be more like a “dynamical system” that is continuously processing and interacting with the environment in real-time while maintaining a stable but ever-changing state.
The philosopher Tim van Gelder used the following example to contrast the computational vs. dynamical descriptions of cognition. Consider the problem of regulating the speed of a steam engine. Steam pressure driving the engine may fluctuate and we need to make sure it turns at a constant speed. The computational approach would involve adding sensors to measure the current speed of the engine, feeding this data into software that represents the engine symbolically with variables and equations, computing the required throttle adjustment, and finally at some later time, signalling the throttle to open or close by a specific amount. This process would then repeat ad infinitum.
The dynamical approach to regulating a steam engine was in fact invented in 1788 by James Watt and is called a centrifugal governor. The engine is connected to a spindle that spins with the engine. The spindle has arms with weights hanging off them such that when the engine speed increases the spindle spins faster causing the weights to move outward due to centrifugal forces thus raising the arms. The arms are connected to the throttle such that as they raise up, the throttle begins to close. Similarly, when the arms move to a lower position (when the engine is spinning slowly) the throttle opens up. The result is an engine that works at a constant speed despite the changes in input energy.
Van Gelder points out that this solution is remarkable because it solves a seemingly computational problem in a non-computational way: there is no flowchart, no procedure or steps to complete, and no symbols. More importantly, the system responds in real-time to changes in the environment without waiting for computations to complete. By using a direct, real-time, continuous feedback loop (the engine controls the position of the arms, and the position of the arms controls the engine) the system maintains stability. Is it then possible that the human brain, a seemingly computational device, can actually be explained in a similarly non-computational way using a dynamical approach?
More:
Plastic Brain
How does the brain learn? Understanding that process would allow one to write software that could learn the same way humans do. But, prior to 1949, no one had a very good answer. It had already been well established that the functional unit of the brain was the neuron, and the structure of these neuronal cells had been studied extensively. However, psychologists did not have a good theory about how neurons produced human behavior.
Enter Donald Hebb, a Canadian Psychologist who was fascinated by how the brain worked. Hebb postulated that neurons form cell assemblies, collections of neurons that act in concert to produce behavior. This idea formed the beginnings of the field of connectionism, an approach to the mind that views complex behavior as emerging from an interconnected network of simpler units. But how do these networks form? To answer this, Hebb proposed a mechanism which has come to be known as “Hebbian Learning.” The idea stated simply is: “Neurons that fire together, wire together.”
The brain is fully connected at birth, but the strength of these connections changes through time as we learn, forming the cell assemblies that Hebb theorized were responsible for behavior. Hebbian learning postulates that when neuron A activates, and then causes neuron B to activate, then the connection strength between the two neurons is increased, and it will be easier for A to activate B in the future. The idea sounds simple, but it goes a long way in explaining how neural networks form in the brain. Not every learning process in the brain can be explained by Hebbian learning, but it does provide an explanation of how complex networks of neurons could form.
After Hebbian learning made its debut in the 1949 book, “The Organization of Behavior,” it then became possible to program computers with the Hebbian rule, giving them the ability to learn. Today, many different types of artificial neural networks (ANNs) are used extensively in the field of artificial intelligence, including applications in face identification, speech and handwriting recognition, financial applications, data mining, and even autonomous vehicles. Hebb’s discovery spawned a whole branch of artificial intelligence and methods for constructing learning mechanisms on computers. ANNs are not yet sufficient for creating human-level intelligence on a computer though. Real neurons are complex biochemical engines, whose behavior can only be approximated with ANNs. Also, human brains come pre-configured to some degree, and without understanding this innate structure well, building large artificial neural networks is not practical.
More:
