Engineering a Sequence Machine Through Spiking Neurons Employing Rank-Order Codes
Sequence memories play an important role in biological systems. For example, the mammalian brain continuously processes, learns and predicts spatio-temporal sequences of sensory inputs. The work described in this dissertation demonstrates how a sequence memory may be built from biologically plausible spiking neural components. The memory is incorporated in a sequence machine, an automaton that can perform on-line learning and prediction of sequences of symbols. The sequence machine comprises an associative memory which is a variant of Pentti Kanerva's Sparse Distributed Memory, together with a separate memory for storing the sequence context or history. The associative memory has at its core a scalable correlation matrix memory employing a localised learning rule which can be implemented with spiking neurons. The symbols constituting a sequence are encoded as rank-ordered N-of-M codes, each code being implemented as a burst of spikes emitted by a layer of neurons. When appropriate neural structures are used the spike bursts maintain coherence and stability as they pass through successive neural layers. The system is modelled using a representation of order that abstracts time, and the abstracted system is shown to perform equivalently to a low-level spiking neural system. The spiking neural implementation of the sequence memory model highlights issues that arise when engineering high-level systems with asynchronous spiking neurons as building blocks. Finally, the sequence learning framework is used to simulate different sequence machine models. The new model proposed here is tested under varied parameters to characterise its performance in terms of the accuracy of its sequence predictions.
The thesis is available as PDF (5.2MB).