Structural synaptic plasticity is an omnipresent mechanism in mammalian brains, involved in learning, memory, and recovery from lesions [1]. Structural plasticity is also a useful computational tool, seeing use in a wide range of applications, including architecture-supervised MNIST handwritten digit classification in the absence of synaptic plasticity rules [2].
Here we show preliminary results of motion selectivity arising from the use of structural plasticity on the SpiNNaker neuromorphic system [3] in real time in conjunction with synaptic plasticity (STDP). The synaptic rewiring rule running on this platform is described in detail in Bogdan et al (2018, [4]). If synapses are allowed to be initialised with random delays then a layer of neurons can perform elementary motion decomposition. Moreover, our proposed architecture allows for unsupervised learning through self-organisation.
Elementary Motion Decomposition through Synaptic Rewiring
The proposed architecture is conducive to neurons becoming sensitised in an unsupervised manner to bars moving in various directions. Bars are encoded using spikes representing "On" and "Off" pixels. The simulations are initialised with no connections and run for around 5 hours. New synapses are formed in two regimes: with random or constant delays. During training a layer is shown either one or multiple movement directions, but during testing they see movement in all directions (in 5 degree increments). The following plots reveal a simplified view of the architecture, a potential setup that could allow for a homeostatic omnidirectional response, an analysis of stability over 20 hours of simulated time and responses for a few neurons in 4 directions. The population response is plotted for each testing direction, while some directions receive extra consideration in a kernel density estimate analysis. Further, individual neuron direction tuning is plotted.
PowerPoint Slide Show mp4 -- a video meant to intuitively reveal a possible solution to the problem of motion detection by using a distribution of delays so that signals arrive at roughly the same time.
testing_instantaneous_firing_rate_angle_of_0 mp4 -- a video showing the spiking activity in the 2D excitatory layer during the testing phase
firing_rate_evolution_with_angle_fill pdf -- evolution of population response when trained from 20 minutes to 40 hours when the network incorporates random delays
firing_rate_evolution_constant_with_angle_fill pdf -- evolution of population response when trained from 40 minutes to 20 hours when the network only incorporates constant delays
Funding
The design and construction of the SpiNNaker machine was supported by EPSRC (the UK Engineering and Physical Sciences Research Council) under grants EP/D07908X/1 and EP/G015740/1, in collaboration with the universities of Southampton, Cambridge and Sheffield and with industry partners ARM Ltd, Silistix Ltd and Thales. Ongoing development of the software is supported by the EU ICT Flagship Human Brain Project (H2020 785907), in collaboration with many university and industry partners across the EU and beyond, and exploration of the capabilities of the machine is supported by the European Research Council under the European Union’s Seventh Framework Programme (FP7/2007-2013) / ERC grant agreement 320689.