Temporal pattern learning and recognition in neural systems

Patterns of sensory information typically change over time, and indeed for some stimuli, such as speech and other auditory information, this is their defining characteristic. One of the most salient differences between conventional artificial systems and their biological counterparts is the way in which timing information is handled. Neurons communicate in an asynchronous fashion using action potentials - electrical spikes that are less than 1 msec duration - but such fine-grained timing is not used in conventional artificial neural networks and machine learning algorithms, which are based on time windows in which the activity of the units corresponds roughly to a time-averaged spiking rate of a neuron or group of neurons. A better understanding of how the brain recognizes and learns distinct sequences of events will be useful in many fields, including speech recognition and bionic ear sound processing.

One research project explores how temporal patterns that change slowly compared with the pace of neural activity, such as sequences of syllables in speech, can be learnt and stored by a neural network. Based on experimental data from the hippocampus, a part of the brain critical to the formation of memories, we have developed a neural network model that explores how sequences of events can be learnt through the adaptation of neural connections as a result of experience. The model demonstrates robust and reliable recognition of sequences of events of varying durations. Other characteristics of learning in the model are currently being investigated.

The neural network is made up of pools of neurons each responding to a particular event. A schematic of a single pool showing its inputs (both external and from other pools) and outputs (to other pools) is shown here. Through adaptation of the connections to other pools, individual neurons learn to respond to specific sequences of events.

Another research project explores how the fine-grained timing structure of the neural response to auditory input can be exploited, with the aim of improving systems for speech recognition. Conventional speech recognition methods rely on spectral information obtained by averaging over tens of milliseconds, but the precise relative timing of action potentials in different auditory nerve fibres conveys information that may be used to separate distinct sources of sound - for example, speech from noise. We seek to understand how this information can be derived from computational models of the peripheral auditory pathway.

These research projects contribute to understanding the neural code and can be used to improve artificial speech processing methods.

People

Dr Sean Byrnes	Prof Tony Burkitt
Dr David Grayden