New publication: "Accelerated Physical Emulation of Bayesian Inference in Spiking Neural Networks"
Our new study "Accelerated Physical Emulation of Bayesian Inference in Spiking Neural Networks" has been published in "Frontiers of Neuroscience - Neuromorphic Engineering". The results demonstrate how the synergy between neuromorphic engineering, machine learning and computational neuroscience can help us discover new types of powerful computational paradigms. A challenge is that these models should be robust against the limitations of neuromorphic hardware. We implemented a probabilistic generative model of spiking neurons on the BrainScaleS "physical model" system with a hierarchical layered structure and applied it to standard image datasets. A learning procedure, using data generated on the hardware, compensated for distorting effects on the substrate. The setup can not only classify images, but is also used for pattern completion and "dreaming" (generating pictures based on the learned dataset) while benefiting from the 10 000-fold acceleration with respect to biological real-time.
The publication is open access and can be found here: https://doi.org/10.3389/fnins.2019.01201