KIP publications

In this work we go back to the original and arguably simplest model known to exhibit critical phenomena, the Ising model for ferromagnetism. Following the model of [Petrovici 2016], we implement an Ising-like spiking neural network that implements the same Boltzmann distribution using recent work on spiking neural networks under Poissonian noise. As such a description is widely used in neuroscience to effectively describe biological models and data one would expect that all the phenomena known from statistical physics can also be observed in these systems.

This work investigates a simplified model, the neural sampling framework introduced by [Buesing 2011], which we modify to include exponentially decaying interactions (resembling biological interactions), capturing the major dynamical difference to LIF sampling. While the global properties, like the unordered phase in the infinite temperature limit and the ordered phase in the zero temperature limit can still be observed. We show that the phase diagram of this model shows richer phenomena than the classical Ising model predicts. For example it allows such a system to converge non-monotonously to its final state even for a non-zero external field.