Skip to content

Cognitive Science 10: How are Cognitive Systems Organized?


Architectures for intelligent agents

There are three types of agent architectures. Simple reflex agents receive sensory stimuli, which goes through a condition-action rule (if-then) and then outputs the action. Goal-based agents have a few more internal steps, including a projection of the effects of the potential actions on the world.Lastly come the learning agents, which have the above abilities, but are also able to learn from their actions when the effects of the actions do not reach a certain standard.

Fodor on the modularity of mind

Fodor proposes a modular mind composed of multiple domain-specific processors, and a central processing unit. The modules are mandatory, which means they act without conscious will. They’re fast, and they are encapsulated- not affected by other modules, unless those provide an output to be processed.

The massive modularity hypothesis

This is in response to theoretical problems with Fodor’s central processor, which I don’t really understand very well. Apparently there is an inherent contradiction. Non-modular processing is context-sensitive, and yet, at least under Fodor’s language of thought hypothesis, the central processor cannot be context-sensitive because it is not informationally encapsulated. Why? I’m not too sure.

So, the proposition is that all information processing is modular, with no central processing going on at all.  A lot of this section applies to Kurzban’s book, so I’ll probably go into more detail here.

Cosmides and Tooby came up with the idea of evolved Darwinian modules as a response to human abilities in processing deontic conditionals. Humans are very good at reasoning regarding obligations, requests, entitlements, but when you abstract out the details, humans become really bad at reasoning similarly. So Cosmides and Tooby proposed that we have evolved certain modules, one of them being one that allows us to detect cheaters really well, because there was evolutionary pressure to do so.

I had wondered while reading Kurzban what specific modules there might be. Bermudez gives a short list:

  • Face recognition
  • Emotion detection
  • Gaze following
  • Folk psychology
  • Intuitive mechanics (folk physics)
  • Folk biology (305).

Research on infants helps shed light on the type of folk mechanics they are predisposed to pick up on. Other evidence for modularity comes from neurological impairments. Face blindness occurs from damage to specific brain region.

Bermudez notes that the information above does not directly contradict the idea of a central processor, and he states that the dominant view in cognitive science is that there is still a domain-general thinking system (306). This helps answer the question of consensus on the modularity hypothesis, although I wish I had some more specific stats.

Cosmides and Tooby give two arguments for the massive modularity hypothesis, based on the assumptions that the human mind is a product of evolution, and that evolution works by natural selection. The point is to try to show that a central processor would not or could not have evolved.

The first is the argument from error. This states that there are too many different areas of fitness/lack of fitness. There would be no way to create a central way to judge when something was fitness promoting. There would be too many systematic errors. So the modules exist to break the problem down into different areas (cheating, mating, relationships, etc) so there will be a closer hugging of the fitness curve.

The argument from statistics says that only a module could internalize a statistical rule for self sacrifice (sacrifice one’s life for 2 sisters, or 1 sister who will reproduce twice as much. . .). A central processor could not have figured such a rule out. But does that rule actually predict people’s altruism?

A response to the above is that there may simply be domain-specific knowledge that humans have innately, since they could not have occurred due to stimuli. There is also the domain-general existence of classical and instrumental conditioning.

A couple stronger responses seek to show the impossibility of the massive modularity hypothesis. First, there is the task of selecting what modules process what information. A cheater detection module would need to have some way of sorting out the information it gets to process. There would need to be a least one module that could process all the information and filter it to where it needed to go. This would make it totally domain-general, which goes against massive modularity.

The second argument looks at behavioral outputs. If many modules recommend different actions, there needs to be some central processing to decide which one wins out.

It looks like Bermudez finds the objections to be more or less convincing, although he does acknowledge the value of much of what massive modularity proponents say.

Leave a Comment

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s

%d bloggers like this: