1. Co-funded by the European Union SlideSP4 Theoretical Neuroscience – HBP 2 nd Periodic Review – June 2016 SP4 Major Achievements: Ramp-Up Phase overall WP4.2:Synaptic Plasticity and Learning - Gruning lab 1 Task 4.2.3 Structures of spiking learning algorithms ALGO MultilayerSpiker (Gruning-lab, Surrey) GIT repository https://fepsweb.eps.surrey.ac.uk/gitlab/HBP 1 SGA1-SP9 neuromorphic SGA1-SP9 neuromorphic SP4 CDP5 ramp-up-phase, work finished SGA1-SP6 NEST Simulator SGA1-SP6 NEST Simulator CDP5 Gardner, Sporea, Gruning. “Learning Spatio- Temporally Encoded Pattern Transformations in Structured Spiking Neural Networks”, Neural Computation 27(12), 2015. Gardner, Gruning, “Optimal Supervised Learning…”. Submitted March 2016.

2. Co-funded by the European Union SlideSP4 Theoretical Neuroscience – HBP 2 nd Periodic Review – June 2016 SP4 Major Achievements: Ramp-Up Phase overall WP4.2:Synaptic Plasticity and Learning - Gruning lab Task 4.2.3 Structures of spiking learning algorithms ALGO MultilayerSpiker (Gruning-lab, Surrey) Big question: How can neural-level mechanisms “conspire” to yield Supervised Learning on systems level with Technically acceptable performance?

3. Co-funded by the European Union SlideSP4 Theoretical Neuroscience – HBP 2 nd Periodic Review – June 2016 SP4 Major Achievements: Ramp-Up Phase overall WP4.2:Synaptic Plasticity and Learning - Gruning lab Task 4.2.3 Structures of spiking learning algorithms ALGO MultilayerSpiker (Gruning-lab, Surrey) Sample Learning Task : Learning a target output spike train in response to a single, fixed input pattern. The network contained n i = 100 input neurons, n h = 10 hidden neurons and a single output neuron. The input pattern was repeatedly presented to the network over 1000 episodes, where each episode lasted T = 500ms. The target output spike train contained five spikes. (A)A spike raster of the input pattern. (B)The activity of one hidden neuron with each episode. (C)The activity of the output, with target output spike times indicated by crosses. (D)An illustration of the multilayer network setup. (E)The evolution of the distance/error between the actual and target output spike trains of the network, given as a moving average of the van-Rossum Distance.

4. Co-funded by the European Union SlideSP4 Theoretical Neuroscience – HBP 2 nd Periodic Review – June 2016 SP4 Major Achievements: Ramp-Up Phase overall WP4.2:Synaptic Plasticity and Learning - Gruning lab Task 4.2.3 Structures of spiking learning algorithms ALGO MultilayerSpiker (Gruning-lab, Surrey) The dependence of the network performance on the number of input patterns, the number of hidden neurons n h, and the number of target output spikes for a single output neuron (Left) Performance as a function of the number of input patterns, for n h = 10 (A), n h = 20 (B), and n h = 30 (C) hidden neurons. In each panel, different curves (blue, red, green) correspond to the number n a =1,5,10 target output spikes identifying each input. (Right) The number of episodes to convergence in learning.

5. Co-funded by the European Union SlideSP4 Theoretical Neuroscience – HBP 2 nd Periodic Review – June 2016 SP4 Major Achievements: Ramp-Up Phase overall WP4.2:Synaptic Plasticity and Learning - Gruning lab Task 4.2.3 Structures of spiking learning algorithms ALGO MultilayerSpiker (Gruning-lab, Surrey) Main Results: 1.New learning algorithm MultiLayerSpiker: Compared to other learning algorithms for spiking neuron networks, we can learn 1.more input-output mappings: 200 pattern pairs here vs 3-4 2.more timed output spikes: up to 10 individually timed spikes here vs 3-5 3.with multiple outputs: up to 30 here vs 1 2.Conclusion for HBP from comparison of diverse supervised learning rules (both biological and machine-learning oriented ones): 1.all need either target or reward signal (“third signals” / three factor rules: neuro- modulators, weak synapses etc) 2.neuromorphic platforms (entire) and NEST (partially) do not yet implement the "infrastructure" to support third signals. 3.=> changes to be implemented in NEST and Spinnaker during SGA1 in CDP 5.