Ensemble of five spiking neural networks trained on MNIST handwritten digits. The time-averaged activity of the hidden and visible layer is shown. The neurons receive no external noise (e.g., from a PRNG) and only acquire their stochasticity due to the background activity of the whole ensemble. Thus, even though the whole system is deterministic, each network is able to stochastically sample from its learned data distribution. Taken from Dold et al. (2018).
A single hierarchical spiking neural network trained on MNIST. Instead of external noise (e.g., Poisson / shot noise), the neurons receive background input from other lateral neurons (red). This allows a completely deterministic network to generate images from the learned data distribution. Taken from Dold et al. (2018).
Ensemble of 15 4-neuron spiking neural networks on the analogue accelerated neuromorphic BrainScaleS system. All networks are trained in parallel to sample from different target distributions, while no external noise is provided to the system. This is a physical implementation of deterministic spiking sampling ensembles with a speed-up of 10,000 compared to biological time. Taken from Dold et al. (2018).
Real-time unsupervised learning of iEEG signals in a rate-based model for error backpropagation in cortical microcircuits, see Senn and Dold et al. (2019)
Spiking neural network completing a linear trajectory (white background) that is slowly uncovered to the network (blue background). We show the time-averaged activity of the visible layer. Using synaptic short-term plasticity helps the network explore possible plausible trajectories ("mixing"). Taken from Zenk (2018).
Spiking neural network completing a non-linear trajectory resulting from a kink in the middle that was not present during training. In addition to synaptic short-term plasticity, the network slowly "forgets" previous parts of the trajectory. Thus, both future and previous parts of the trajectory have to be reconstructed by the network. The effective input the network sees is shown on the bottom. Taken from Zenk (2018).