Encoding and Retrieval Efficiency of Episodic Data in a Modified Sparse Distributed Memory System

Encoding and Retrieval Efficiency of Episodic Data in a Modified Sparse Distributed Memory System Sidney K. D’Mello (sdmello@memphis.edu) Computer Science Department & The Institute for Intelligent Systems, 365 Innovation Drive, Memphis, TN 38152, USA Uma Ramamurthy (urmmrthy@memphis.edu ) The Institute for Intelligent Systems, 365 Innovation Drive Memphis, TN 38152, USA Stan Franklin (franklin@memphis.edu ) Computer Science Department & The Institute for Intelligent Systems, 365 Innovation Drive, Memphis, TN 38152, USA episodic memory, and autobiographical/declarative memory (Ramamurthy, D’Mello, & Franklin, 2003). We hypothesize that information stored in TEM, which has not decayed away, is consolidated into declarative memory at certain intervals. Transient episodic and declarative memories have distributed representations in IDA. There is evidence that this is also the case in animal nervous systems. Some of these memory models are motivated by Sparse Distributed Memory (Kanerva, 1988). This is reasonable due to several similarities between SDM and human memory systems such as knowing that one knows, tip-of-the-tongue effect, rehearsal, momentary feelings of familiarity, and interference. The focus of this paper is on a modified SDM architecture that promises to be a good candidate for use as a TEM in software agents such as IDA. Abstract This paper presents detailed simulation results on a modified Sparse Distributed Memory (SDM) system. We have modified Kanerva’s original SDM system into an architecture with a ternary memory space. This enables the memory to be used as a Transient Episodic Memory (TEM) in cognitive software agents. TEM is a memory with high specificity and low retention, used for events having features of a particular time and place. Our earlier work focused on perfunctory, proof of concept assessments on the modified SDM system. This paper presents a detailed experimental evaluation of the modified SDM system with regard to its ability to store and retrieve episodic information. Introduction Episodic memory is for events having features of a particular time and place (Baddeley et al, 2001). This memory system is associative in nature and content- addressable. It has been proposed that working memory probably includes an episodic buffer that can hold episodic information for a short duration (Baddeley, 2000). Humans have a content-addressable, associative, transient episodic memory with a decay rate measured in hours (Conway, 2001). Humans are able to recall in great detail events of the current day – where they park their cars, whom they met that morning, what they discussed, what they had for meals, etc. These details of the events/episodes stay with us only for short durations – for some hours. We hypothesize that for cognitive agents to recall such details of episodes while they interact with and adapt to their dynamic environments, they need a transient episodic memory (TEM). The Intelligent Distribution Agent (IDA) is one such cognitive software agent endowed with a TEM (Baars, & Franklin. 2003, Franklin et al in review). IDA is a cognitive software agent (Franklin, 1997) developed for the U.S. Navy. At the end of each sailor’s tour of duty, he or she is assigned to a new billet by a person called a detailer. IDA’s task is to facilitate this process by completely automating the role of a detailer. The IDA technology (Franklin, 2001) has a number of different memory systems, including working memory, transient Sparse Distributed Memory SDM implements a content-addressable random access memory. Its address space is in the order of 2 1000 . Of this space, you choose a manageable, uniform random sample, say 2 20 , of allowable locations. These are called hard locations. Thus the hard locations are sparse in this address space. Many hard locations participate in storing and retrieving of any datum, resulting in the distributed nature of this architecture. Hamming distance is used to measure the distance between any two points in this memory space. Each hard location is a bit vector of length 1000, storing data in 1000 counters with a range of -40 to 40. Each datum to be written to SDM is a bit vector of length 1000. Writing 1 to a counter results in incrementing the counter, while writing a 0 decrements the counter. To write in this memory architecture, you select an access sphere centered at location X. So, to write a datum to X, you simply write to all the hard locations (typically 1000 of them) within X’s access sphere. This results in distributed storage. This also naturally provides for memory rehearsal – a memory trace being rehearsed can be written many times and each time to about 1000 locations. Similar to writing, retrieving from SDM involves the same concept of access sphere – you read all the hard