Do brain cells need to be connected to have meaning?
The classic theory of the brain is one of connections, in which the brain consists of a network of neurons that interact with each other to allow us to think, see, interpret, and understand the world around us. In this model, called distributed representation, an individual neuron by itself has no inherent meaning, but only contributes to a pattern of neuronal activity that has meaning. For example, a certain pattern of many neurons fires when you think “dog” and another pattern for “cat.”
"The belief in distributed representation theory is that a concept or object is not represented by a single neuron in the brain but by a pattern of activations over a number of neurons," explains Asim Roy, a professor of information systems at Arizona State University, to Medical Xpress . "Thus there is no single neuron in the brain representing a cat or a dog. Proponents of this theory claim that a cat or a dog is represented by its microfeatures such as legs, ears, body, tail, and so on. However, they think that neurons have absolutely no meaning on a stand-alone basis. Therefore, they go further and claim that these microfeatures are at the subsymbolic level, which means that meaning arises only when you consider the pattern of activations as a whole. Therefore, there are no neurons representing legs, ears, body, tail, etc. The representation is at a much lower level."
Roy is among a number of scientists working in the fields of neuroscience and artificial intelligence (AI) who suspect that the brain may not be as connected as distributed representation suggests. The basis of their alternative model, called localist representation, is that a single neuron can represent a dog, a cat, or any other object or concept. These neurons can be considered symbols since they have meaning on a stand-alone basis. However, as Roy explains, this doesn’t necessarily mean only one neuron represents a dog; such “concept cells” are high-level neurons, which fire in response to the firing of an assortment of low-level neurons that represent the legs, ears, body, tail, etc.
"In localist representation, there could be separate neurons for a dog and a cat, and also neurons for legs, ears, body, tail, etc.," he said. "It’s very similar to the model in my paper for word recognition, which is an old model from James McClelland [Chair of the Psychology Department at Stanford University] and [the late pioneering neuroscientist] David Rumelhart. You have low-level neurons that detect letters of the alphabet and then high-level neurons for individual words. So letter neurons and word neurons, they both exist."
The origins of this dispute between localist and distributed representation goes back to the early ’80s, to a dispute between the symbol processing hypothesis of artificial intelligence (AI) and the subsymbolic paradigm of connectionists. In the past 30 years, the debate has only intensified.
