# KIP publications

year | 2017 |

author(s) | A. Baumbach, M. A. Petrovici, L. Leng, O. J. Breitwieser, D. Stoeckel, I. Bytschok, J. Schemmel, K. Meier |

title | Bayesian computing with spikes |

KIP-Nummer | HD-KIP 17-138 |

KIP-Gruppe(n) | F9 |

document type | Paper |

Keywords | neuromorphic computing, non-linear systems, statistical physics |

source | 1st HBP Student Conference |

Abstract (en) | In view of mounting evidence that brains solve various cognitive tasks by performing Bayesian inference [1, 2], theoretical models have been developed to explain how such computations can be implemented in networks of spiking neurons [3, 4, 5, 6]. In the neural sampling framework [3, 4], the activity of spiking neurons is interpreted as sampling from an underlying probability distribution that is shaped by the network’s parameters. Operating under this premise, the models developed in [4, 5, 6] have shown how biological neural networks can make use of the high-conductance state to achieve equivalent computational capabilities. This allows, for example, the straightforward reproduction of well-known stochastic behavioral phenomena such as perceptual ambiguity with biologically plausible spiking neural networks [5]. Here, we briefly review several applications of such spiking sampling networks, including a somewhat exotic discussion of their connection to statistical physics. The computational abilities of these networks can be directly put to use in classical machine learning tasks such as handwritten digit recognition. In addition to achieving a similar performance when compared to traditional approaches such as Gibbs sampling, our spiking networks can profit from biology-inspired features such as short-term plasticity to simultaneously provide good generative capabilities, which are otherwise difficult to achieve [7]. Owing to the fact that our networks employ LIF neurons - a de facto standard model for neuromorphic devices - our networks can directly benefit from the advantages offered by these substrates. In particular, the emulation on an accelerated mixed-signal chip [8] has recently been demonstrated [9], which paves the way for larger-scale, accelerated applications for demanding computational tasks [7, 10]. The approximate equivalence of the dynamics of LIF networks to the so-called Glauber dynamics in microscopic models of magnetic materials [11] raises the question whether macroscopic effects, such as the relationship between magnetization, external field and temperature, are also conserved. Preliminary experiments suggest consistent behavior in the perturbative limit (large temperatures, weak external excitation), but deviant behavior around critical points. |

Datei |