Generalization in probabilistic RAM nets

Clarkson, T.G.; Guan, Y.; Taylor, John G.; Gorse, Denise

doi:10.1109/72.207603

Cited by 17 publications

(8 citation statements)

References 6 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Yet finer details of the system's behaviour can be obtained from the differential equation (5). Since The definition of two meaningful terms helps the interpretation of the above equation.…”

Section: Temporal Evolution Of Learningmentioning

confidence: 99%

“…The first step is achieved by making the differential equation (5) , which describes the process, equal to zero. The second step is to find a suitable positive definite function which satisfies the Liapunov conditions for asymptotic global stability [7].…”

Section: Stability Analysismentioning

confidence: 99%

“…The assumption of a very small learning rate p allows for the continuous approximation of the difference equation [8] above turning it into a differential equation (5). The time derivative of the average firing probability (&) of the addressed site (state) is the basis for most of the analysis to follow.…”

Section: R Amentioning

confidence: 99%

“…Random noise has been added to the PRAM reinforcement learning [5] to endow the neuron (look-up table) with generalisation power. Noisy training had been devised [I41 for better exploring the environment and acquiring more information about it.…”

Section: Introductionmentioning

confidence: 99%

“…It is the first part of a modular approach to analyse the formation of basins of attraction by PRAM noisy reinforcement learning [ll] and its generalisation [5]. Instead of making the analysis considering the simultaneous training of all sites, the problem of training the neuron is decomposed into training a single site and then accounting for the effect of the different input patterns and the noise added to them on all sites.…”

Section: Introductionmentioning

confidence: 99%

See 4 more Smart Citations

Stability analysis of pRAM reinforcement learning

Adeodato

Taylor²

Proceedings 4th Brazilian Symposium on Neural Networks

Self Cite

View full text Add to dashboard Cite

Generalisation has been a major issue in RAM-based neural networks. In PRAM networks generalisation is produced by noisy reinforcement learning; a completely hardware implementable (built-in) algorithm. This paper presents the first part of a modular technique t o analyse the formation of the basins of attraction i n such systems. It proves that reinforcement learning i n a single P R A M site is a globally stable system i n the continuous limit of incremental learning. It also shows how the stable state depends on the penalty/reward ratio and on the learning rate. The evolution of learning in the tame domain shows the effects of the initial state and of the halting moment in the final state. The paper ends with considerations on how noise contributes to the formation of basins of attraction i n P R A M neurons.

show abstract

Section: Temporal Evolution Of Learningmentioning

confidence: 99%

Section: Stability Analysismentioning

confidence: 99%

Section: R Amentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations