Neuromorphic implementations of polychronous spiking neural networks

The object of this thesis is to investigate polychronous spiking neural networks using neuromorphic implementations. This type of neural network has enormous memory capacity as it can store far more spatio-temporal patterns than it has neurons, which could help to explain how the human cortex can have such a diversity of behaviour with a mere 1011 neurons. To date, most of the published polychronous spiking neural networks have been implemented using software neuron models and such simulations are not capable of achieving emulation of large-scale neural networks in real time. We therefore present a mixed-signal implementation of a reconfigurable, polychronous spiking neural network with a vast capacity for storing spatio-temporal patterns. The work presented in this thesis includes the design of a polychronous spiking neural network using a novel delay-adaptation algorithm, an FPGA implementation of the proposed neural network, an analogue implementation of the proposed neural network, and their integration into a mixed-signal platform. Rather than using a weight-adaptation algorithm such as Spike Timing Dependent Plasticity (STDP) to prune and select appropriate subsets of delays during training, the proposed neural network generates delay paths de novo, so that only connections that actually appear in the training patterns will be created. This allows the proposed network to use all the axons (variables) to store information. Spike Timing Dependent Delay Plasticity (STDDP) is proposed to fine-tune the delays of the axons and add dynamics to the network. The FPGA implementation uses a time-multiplexing approach allowing us to achieve 4096 (4k) neurons and up to 1.15 million programmable delay axons. The analogue implementation comprises 50 neurons and 400 axons. Compared to the digital implementation, the analogue implementation is more biological plausible as the computation in biological neurons is conducted with analogue variables. An analogue memory with a novel structure and a very low leakage rate was designed and characterised for the analogue axon. In the mixed-signal platform, 4K analogue neurons were achieved by using a time-multiplexing approach. Testing results show that the mixed-signal implementation of the proposed neural network is capable of successfully recalling up to 96% of the stored patterns. The results also show that the neural network is robust to noise and problems such as device mismatch and process variation