Since the course is called Computational Models of Cognitive Processes, it’s interesting to note that there are some definite, intrinsic similarities between the human thought processes and mathematical and computational models that stand out to me. Even though the course is just beginning I am getting a sense that probabilities play a large role in both computational models and the way our brains work. We often do things based on how likely it is to succeed or how many times a given action has been performed, and our proficiency at it.
Computer models can behave the same way. Probabilistic— and even neural— networks, are a definitive corner stone of artificial intelligence. These models can ‘learn’ certain information and ‘act’ according to certain probabilities that can either be ‘trained’ or manually entered. For example, Hofstadter discusses a hypothetical machine that has probabilistic properties on page 58 of his book. This machine picks apart sequences of numbers and searches for patterns, but he doesn’t want to approach it in a standard engineering sort of way (brute-forcing the sequence computationally). Instead he intends to design a machine that functions in the same way as the human brain. This requires creativity and finesse. Looking at how humans solve problems, he considers a type of ‘perceptual glue’ that makes certain sequences stand out to us:
Given this number sequence, what stands out to you?
1924522222123456789010101
Notably, the presence of the repeating 2’s, or even the sequence from one to nine, in order, somewhere in the middle. However, if you were really paying attention you would have also noticed the repeating sequence of 01’s. This is what Hofstadter calls ‘perceptual glue’. Our perception sticks these things together for us. Our brains notice these sub-sequences first because they have a higher probability of being important to us in some way. In a similar fashion we can, and Hofstadter attempts to, design computer models to look for these groupings first and, based on the probability that they are important (if they appear more than once, etc.), work from there.
In light of this opportunity to design a machine which works in a ‘cognitive’ way he manages several problems at once. First off, the machine is more flexible and efficient, it won’t always approach a problem in a sequential manner; instead it can find sections that are ‘perceptually glued’ together and based on the likely-hood of success it will pursue that path. If we, as humans, inversely, were to attempt every problem sequentially as if from an ordered list of procedures the same way, every time, there would arise problems that might end up unsolvable to us. People don’t behave like this, and ‘intelligent’ machines wouldn’t either. A machine that follows a list of orders every time can hardly be called artificial intelligence, can it?
-James
