Week 4 - Cognitive Architecture

• Just like last week, this week we focused on understanding cognitive architecture. To summarize the discussion, cognitive architecture is a theory and we can claim that a system has cognitive architecture if it can behave like a human mind.

• The article considers learning to be on a continuum with novice learners being on the left side of the spectrum and experts on the right side.

• When we presented with a new set of information, it first goes in our working memory which is short term and has relatively smaller storage capacity. Adding more information in the working memory causes cognitive overload which results in low level learning/understanding of the concepts.

• In order to truly understand how ‘learning’ and ‘understanding’ takes place and how a learner can shift from the left side of the continuum to the right side, we need to look at the level of interactivity. Whenever we are introduced with a new set of information, if it has high interactive elements then the working memory will not be able to process it properly. The ideal situation would be to look at the information in isolated and non-interactive manner first.

• Schemas were also discussed at length because it is due to schemas the cognitive load on the working memory is decreased. If there are proper schemas built then the mind can process information rapidly in the working memory even if the information has high interactive elements.

• Transfer of knowledge is also very important because if a learner is able to efficiently transfer his learning in different settings, then it shows that there are central executives formed in his/her long term memory.

• If the information is new, then there will be fewer schemas available in the memory due to which a learner might find it difficult to understand the concept.

• Expertise is dependent upon automation, which implies that a learner can process information without causing cognitive overload in the working memory.

Session 5

This week we learnt about cognitive architecture. Understanding Cognitive architecture (CA) is imperative because it can lead to designing an artificial intelligent system that can replicate human capabilities. CA is more about the underlying infrastructure of an intelligent system. It is not concerned with what is stored in the memory because the content in the memory can change over time and CA deals with those aspects that are constant overtime. CA varies from expert systems as the latter have finite information based on which the possible outcomes are also finite.

There are a few examples of how CA is utilized to construct machines that cognitive abilities. One of them is ACT-R which tries to model human behavior. It has set modules (sensory, motor, intentional and declarative) which process different type of information. In order to understand better how ACT-R works, please follow this link as it shows how a small robot tries to find its way out of a maze (ACT-R Robot). Soar is another example. It is slightly more advanced than ACT-R because it has episodic and semantic memories. It is more goal oriented and in order to achieve it, it dynamically selects the best possible method/way/technique from the information it already has to get to the end result. If it does not have prior information about certain goals then it tries to collect that information first. To truly understand how a Soar system functions, please follow this link as it depicts an example of a train track and train engine (Soar Explained). The other examples of CA would be ICARUS and PRODIGY.

CAs have capabilities such as recognizing, categorizing, decision making, situation assessment, prediction, planning etc which enables them to function dynamically. In addition, the properties of CA include representation, organization, utilization and refinement of knowledge. When evaluating CA aspects such as its generality, taskability, efficacy, reactivity, improvability and autonomy are taken into consideration. The most important take away from this is to understand that CA does not follow “all or none” approach implying that there is possibility that an intelligent agent might not have CA related capabilities. Similarly, if it does not have all the capabilities then it cannot be expected to have all the properties CA has to offer and cannot be assessed or evaluated against all the evaluation criteria mentioned above. Systems based on CA are dynamic and hence will differ in properties.