Hirase, Hajime; (1997) Performance analysis of a recurrently connected auto-associative memory. Doctoral thesis (Ph.D.), University College London (United Kingdom).

Text Performance_analysis_of_a_recu.pdf Download (4MB)

Abstract

Associative networks have long been regarded as a biologically plausible mechanism for memory storage and retrieval. When the input to the associative network includes erroneous information or the network is incompletely connected, the performance critically depends on the thresholding strategy. Applying a low threshold increases the probability that spurious cell will fire. On the other hand, a high threshold decreases the probability that correct cells will fire. This thesis is concerned with finding improved thresholding strategies in partially connected recurrent associative networks that have binary weights. In the case of simple feedforward associative networks, the choice of optimal threshold can be analytically computed on the basis of statistical expectations. However, when the network is recurrently connected, such a mathematical analysis of the network behaviour becomes rather complex. Using the formalism proposed by Gibson and Robinson (1992), optimal thresholding sequences are identified for a partially connected recurrent associative network. Surprisingly, the value of the threshold is found to be proportional to the current activation of the network. Based on this result, it is concluded that a linear thresholding strategy (where the level of global threshold is computed linearly from the current activation of the network) is near optimal as far as storage capacity (but not retrieval speed) is concerned. This result is validated in a simulation study. Also, the coefficients of the linear threshold equation were found to be independent of the memory loading. The linear thresholding strategy is a global thresholding strategy in that every cell undergoes the same threshold. It is argued that a cell-specific thresholding, in which each cell receives a different level of threshold, produces a better recall performance. A new method to implement local threshold using interneuron learning is proposed. In this paradigm, real-valued synapses are introduced to the inhibitory interneuron that effectively set the threshold level of each principal cell. By varying the inhibitory synaptic efficacy, local thresholding is achieved. It is demonstrated that a simple Hebbian mechanism in the projection from the interneuron can significantly improve the performance of multi-step recall. Finally, implications for a biological form of associative memory are discussed especially in relation to the hippocampus, which is thought to be involved in the process of intermediate memory and in consolidation.

Download activity - last month

Export as JSONCSVXML

Download activity - last 12 months

Export as XMLCSVJSON

Downloads by country - last 12 months

Export as CSVJSONXML

Archive Staff Only

View Item