Simulating synaptic rewiring on SpiNNaker
poster
posted on 2018-03-27, 09:02 authored by Petruț Antoniu BogdanPetruț Antoniu Bogdan, Steve FurberStructural synaptic plasticity is an omnipresent mechanism in mammalian brains, involved in learning, memory, and recovery from lesions. Structural plasticity is also a useful computational tool, used in automatically generating connectivity based on experimental activity data, exploring network states for Bayesian inference and assisting wider-spread synaptic plasticity rules for better performance.
The structural organisation of cortical areas is not random; topographic maps are common-place in sensory processing centres. Topographic organisation allows optimal wiring between neurons, multimodal sensory integration, and performs input dimensionality reduction. We have designed an efficient framework which can be used to simulate models of structural plasticity on the SpiNNaker neuromorphic system, in real time, in conjunction with synaptic plasticity rules, such as spike-timing dependent plasticity (STDP). A model of generic topographic map formation is implemented, using our framework, making use of both activity-dependent and independent processes. In agreement with the work by Bamford et al. (2010), we show that structural plasticity in the form of synaptic rewiring refines an initially rough topographic map and embeds input preferences into the network connectivity. Additionally, it can also be used to generate topographic maps between layers of neurons with minimal initial connectivity, and stabilize projections which would otherwise be unstable.
Finally, we show that supervised MNIST handwritten digit classification can be performed in the absence of synaptic plasticity rules (i.e. rules which change the weights or efficacies of connections). This is not a state-of-the-art MNIST classification network (it achieves a modest accuracy of 78% and an RMSE of 2.01 with non-filtered inputs, performance drops when filtered inputs are used: an accuracy of 71% and an RMSE of 2.38) as each input digit class is represented only as an average for that class, but it serves here to demonstrate that synaptic rewiring can enable a network to learn, unsupervised, the statistics of its inputs. Moreover, with the current network and input configuration, the quality of the classification is critically dependent on the sampling mechanism employed in the formation of new synapses. Random rewiring, as opposed to preferentially forming connections to neurons that have spiked recently, could achieve accurate classification only if operating in conjunction with STDP or some other mechanism to prevent the pruning of useful synapses.
The structural organisation of cortical areas is not random; topographic maps are common-place in sensory processing centres. Topographic organisation allows optimal wiring between neurons, multimodal sensory integration, and performs input dimensionality reduction. We have designed an efficient framework which can be used to simulate models of structural plasticity on the SpiNNaker neuromorphic system, in real time, in conjunction with synaptic plasticity rules, such as spike-timing dependent plasticity (STDP). A model of generic topographic map formation is implemented, using our framework, making use of both activity-dependent and independent processes. In agreement with the work by Bamford et al. (2010), we show that structural plasticity in the form of synaptic rewiring refines an initially rough topographic map and embeds input preferences into the network connectivity. Additionally, it can also be used to generate topographic maps between layers of neurons with minimal initial connectivity, and stabilize projections which would otherwise be unstable.
Finally, we show that supervised MNIST handwritten digit classification can be performed in the absence of synaptic plasticity rules (i.e. rules which change the weights or efficacies of connections). This is not a state-of-the-art MNIST classification network (it achieves a modest accuracy of 78% and an RMSE of 2.01 with non-filtered inputs, performance drops when filtered inputs are used: an accuracy of 71% and an RMSE of 2.38) as each input digit class is represented only as an average for that class, but it serves here to demonstrate that synaptic rewiring can enable a network to learn, unsupervised, the statistics of its inputs. Moreover, with the current network and input configuration, the quality of the classification is critically dependent on the sampling mechanism employed in the formation of new synapses. Random rewiring, as opposed to preferentially forming connections to neurons that have spiked recently, could achieve accurate classification only if operating in conjunction with STDP or some other mechanism to prevent the pruning of useful synapses.
