Another way in which the brain differs from a standard serial computer is that processing doesn't just go in only one direction at a time. Thus, not only are lots of things happening at the same time (parallelism), but they are also going both forward and backward too. This is known as interactivity, or recurrence, or bidirectional connectivity. Think of the brain as having hierarchically organized processing areas, so that visual stimuli, for example, are first processed in a very simple, low-level way (e.g., in terms of the little oriented lines present in the image), and then in subsequent stages more sophisticated features are represented (combinations of lines, parts, objects, configurations of objects, etc.). This is at least approximately correct. In such a system, interactivity amounts to simultaneous bottom-up and top-down processing, where information flows from the simple to the more complex, and also from the more complex down to the simple. When combined with parallelism and gradedness, interactivity leads to a satisfying solution to a number of otherwise perplexing phenomena.
For example, it was well documented by the 1970s that people are faster and more accurate at identifying letters in the context of words than in the context of random letters (the word superiority effect). This finding was perplexing from the unidirectional serial computer perspective: Letters must be identified before words can be read, so how could the context of a word help in the identification of a letter? However, the finding seems natural within an interactive processing perspective: Information from the higher word level can come back down and affect processing at the lower letter level. Gradedness is critical here too, because it allows weak, first-guess estimates at the letter level to go up and activate a first-guess at the word level, which then comes back down and resonates with the first-guess letter estimates to home in on the overall representation of the word and its letters. This explanation of the word superiority effect was proposed by McClellandRumelhart81. Thus, interactivity is important for the bootstrapping and multiple constraint satisfaction processes described earlier, because it allows constraints from all levels of processing to be used to bootstrap and converge on a good overall solution.
Figure: Ambiguous letters can be disambiguated in the context of words (Selfridge, 1955), demonstrating interactivity between word-level processing and letter-level processing.
There are numerous other examples of interactivity in the psychological literature, many of which involve stimuli that are ambiguous in isolation, but not in context. A classic example is shown in figure 1.6, where the words constrain an ambiguous stimulus to look more like an H in one case and an A in the other.