Biologically Inspired Cognitive Architectures

Posted: October 28, 2011 at 12:42 pm

On one hand we have the deep machine learning systems discussed in the previous post, which are quite complex (from a mathematical perspective) and only roughly informed by neuro and cognitive sciences. On the other hand are cognitive architectures that are inspired by biological knowledge (BICAs). These are systems that attempt to balance high level (psychological knowledge) with low level (neurological knowledge) in a tractable system that models mind-brains. These systems can be purely symbolic, connectionist (sub-symbolic) or hybrids of the two. While they are biologically inspired they favour a high level abstraction of cognition over fine details of brain (neuron) function.

Symbolic BICAs follow the “physical symbol system” hypothesis, where cognitive functions can be considered symbolic, and therefore realizable independently of the brain, in any sufficiently complete symbolic system (like a computer). These systems emphasize centralized control and reasoning over emergence and decentralized processes. These systems (In particular ACT-R and CLARION) have already been rejected for this project as their emphasis on reasoning and task-oriented problem solving is likely too restrictive to describe dream cognition.

Connectionist (emergentist) architectures reject the high level and centralized abstractions of cognition for a network of similar neuron-like units that operate on a “sub-symbolic” level. These systems are sub-symbolic in that information is encoded in complex states that are not easily converted into symbolic representation. This is the key: symbolic systems ignore complex and emergent aspects of complex networks of units in order to emphasize the high level and symbolic representation of cognitive states. Connectionist systems apparently have no symbolic level at all (it is implicit) and depend entirely on the complex interaction between simple components.

A connectionist approach (the Self-Organizing Map) was used in Dreaming Machine #1 and #2. One of the major plans in Dreaming Machine #3 is that the system develops a conceptual system, that is one with representations that could be interpreted as symbolic. Even if these representations are simply patterns of activation of neuron-like units, they are expected to be abstracted away from low level sensory representations. There are also developmentally oriented architectures that will not be discussed as a strong view of the developmental account of conceptual development has been ruled out.

Another option are hybrid architectures that combine various approaches from above. Of particular interest for this project are connectionist – symbolic systems as they must explicitly deal with the relation between low level emergent representations, and the symbolic operation on those representations. While the role of problem solving and reasoning is unclear in dreaming, these architectures may provide a more rounded conception of cognition that does not abstract away the low level complexity that is likely crucial to dream cognition.

I have been reading about these cognitive architectures for some time, but have never been able to map the theories of dreaming I’ve been reading to their structure, especially those symbolic architectures. The LIDA (Learing Intelligent Distribution Agent) is the first architecture I have come across that I could see immediate relevance to this project. Major features of this architecture that make it attractive include:

(1) The whole system is based on a “cognitive cycle” that relates sensory impressions with memory, and eventually with actions in the world. This is roughly similar to the cycle Dreaming Machine #1 and #2 operate on: Stimulus → Memory → Output

(2) Perception is considered a complex and multi-layered process that includes interaction with memory and corresponds to the constructive aspects of perception.

(3) Structures in working memory have a decay in their activation, so even when new sensory data arrives, previous structures related to previous sensory data are still somewhat active. This corresponds to latent activation of perceptual systems, which, for this project, is how dreams are initiated and why they relate to waking experience.

(4) LIDA “makes sense” of stimuli by considering it in relation to previous memories while it constructs a “model”.

(5) Built-in mechanisms for attention constrain higher level processes (avoiding combinatorial explosion), and based on the activation of precepts. The sum  or average of activation of percepts could be considered a measure of degree of stimulation (useful in triggering a sleep state, or causing waking).

(6) A built-in notion of consciousness, although I do not understand it at this point.

(7) LIDA processes are mapped to neural and cognitive processes in a highly complete fashion.

With all these features in mind its important to note that LIDA is still oriented toward the intentional and goal-direction action of an agent in an environment. One possibility is to use a subset of the architecture, as it pertains to dreaming. A further option would be to create an additional module for dreaming that could effect memory consolidation and perhaps creativity. Unfortunately, the official LIDA software framework is JAVA only, which I am not very well versed in, and could cause difficulties when integrating with other libraries, such as openFrameworks, OpenCV, etc..

LIDA certainly requires a deeper look, and perhaps simplify the dreaming machine system by using an existing cognitive framework. It is the first architecture that appears to already fit into my thinking on dream cognition. Until this point I could not see how these architectures would relate.