Bitstream neurons for graph colouring
Peter Burge, John S. Shawe-Taylor · ePrints Soton (University of Southampton) · 1996
We introduce the Multi State Bitstream Neuron. By replacing the stochastic activation function with stochastic weights the MSBSN is shown to approximate a Generalised Boltzmann Machine. Benchmarks show the algorithm performs as well as the Boltzmann algorithm whilst the MSBSN lends itself to a very compact and fast hardware implementation. 1 Introduction In [1] Shawe-Taylor and Zerovnik introduced the Generalised Boltzmann Machine (GBM) as an extension to the Boltzmann machine that enables us to map constraint problems, requiring more than two states, onto a recurrent neural network. In [2] experiments were performed using the Mean Field Annealing approach to graph colouring using the Petford and Welsh algorithm [3] as a GBM. In this paper, we extend the use of bitstreams from the bi-polar, stochastically connected Boltzmann machine [4] to a recurrent network of stochastically connected multi-state bit stream neurons (MSBSN). We compare the performance of the resultant network with t...