Biomedical Engineering - Theses
Permanent URI for this collection
Search Results
1 - 2 of 2
-
ItemNeural network models of pitch perception in normal and implanted ears( 2015)Pitch is the perceptual correlate of sound frequency and is important for using speech prosody, understanding tonal languages, and music appreciation. According to pitch perception theories, pitch is encoded by the place along the cochlea that has the maximum rate of excitation and the timing of the neural activity caused by cochlear excitation. As one of the most successful neural prosthesis, cochlear implants (CIs) have enabled most recipients to achieve good speech perception in favourable listening conditions. Perceiving a precise pitch, however, is still a challenge for many CI users. The goal of this study is to develop computational models of normal and CI hearing to advance our understanding of the mechanisms of pitch perception in both cases and to investigate the factors that affect pitch perception in electrical hearing. Based on the two possible neural codes for pitch perception, two models of pitch perception were developed: the place model and the integrated model. The models simulated a common psychophysical experiment usually referred to as pitch ranking − where subjects are asked to decide which of the two sequentially presented sounds has a higher pitch – by receiving two stimuli and generating two outputs, the higher-amplitude of which would indicate the higher-pitched stimulus. Synthesised vowels with defined pitches were the sound stimuli used in this study. An artificial neural network (ANN) constituted the core of the models. Inputs to the ANN were place pitch information for the place model and both place and temporal pitch information for the integrated model. Applying the error back-propagation algorithm, the ANN was trained on a training set of pitch pairs. The performance of the pitch perception model was measured using a previously unseen test set of pitch pairs. Place code for pitch perception was extracted from simulated auditory peripheral outputs by averaging the rate of activity in the auditory nerve (AN) over time. An acoustical and an electrical model of the auditory periphery were used to simulate the activity of the AN in normal and CI hearing, respectively. The activity of the AN was further processed through a spiking neural network (SNN) to extract the temporal code of pitch. Synaptic connections in the SNN were modified by spike-timing-dependent-plasticity (STDP) to generate pitch-related precisely-timed neural activities in the SNN output neurons. Pooled inter-spike-interval histogram (ISIH) across the SNN output neurons was found to be indicative of pitch. Validation of both pitch perception models was performed by comparing their performance with psychophysical results. The place model was applied to investigate the impact of stimulation field spread on pitch perception in CI hearing using two commercial sound processing strategies. Simulation results showed that 1) the model could replicate the performance of normal and average-performing CI listeners and 2) providing focused stimulation fields in CI hearing can be beneficial, depending on the type of sound processing strategy. The integrated model was used to explore the role of and interaction between place and temporal cues in performing simulated pitch ranking tasks. Simulation results associated with the integrated model revealed that 1) temporal cues for pitch perception compensated for missing place cues in listening conditions such as a telephone conversation where low-frequency content of the signal was suppressed and 2) new strategies with improved temporal information can improve pitch perception in CI hearing, provided temporal and place information are consistent. Although drawing general conclusions about auditory perception would eventually require psychophysical experiments, computational models of auditory perception such as this work assists in focusing human testing upon factors that demonstrate the strongest impact on the auditory performance of normal and CI listeners. This would therefore lead to the more rapid development of CI sound processing strategies.
-
ItemMathematical modeling of brain networks: from synaptic plasticity to behavior( 2014)A fundamental goal of neuroscience is to understand how the brain encodes and processes information and how the networks and structures involved are formed. In this thesis, we use theoretical approaches to further our understanding of brain function. First, we investigate how experimentally-based learning rules lead to the formation of different network structures, through unsupervised learning. Second, we investigate how different experimentally-based neural models and network structures enable different types of information processing, such as goal-directed, top-down processing. Third, we consider how reinforcement learning arising from synaptic plasticity mechanisms can coexist with unsupervised learning during the operant conditioning of neural firing rates. The unsupervised learning rule spiking-timing-dependent plasticity (STDP) has been shown to selectively potentiate feed-forward connections with specific axonal delays, enabling functions such as sound localization in the auditory brainstem of the barn owl. We demonstrate a similar selective potentiation for the recurrent connections in a network with axonal delays corresponding to the period of incoming oscillatory activity with frequencies in the range of 100-300Hz. For lower frequency oscillations, such as gamma (60Hz), we show that multiple, recurrently connected groups of neurons could encode not only the oscillation frequency but also a time lag between different sets of oscillations. These results have the potential to help explain missing fundamental pitch perception in the auditory brainstem and the formation of neuronal ensembles (or cell assemblies) in the cortex, respectively. Neural systems are able to perform top-down processing of stimulus information and flexibly select behaviors appropriate to the environment and present goals. Based upon previous experimental and theoretical studies, we propose that information in higher-level areas of the cortex, such as the prefrontal cortex, is encoded in the amplitude and phase of neural oscillations, such as gamma oscillations, and that this activity is gated by two mechanisms: top-down feedback and coherence between these oscillations. By forming these units into circuits that can perform logic operations, we identify the different ways in which operations can be initiated and manipulated by top-down feedback. We demonstrate that more sophisticated and flexible top-down control is possible when the gain of units is modulated by two mechanisms. We explore how different network properties affect top-down control and make predictions about the likely connectivities between certain brain regions. Typical and well-studied examples of behavioral learning are those in which the firing rates of individual cortical neurons in monkeys are increased using rewards. These results have been reproduced using reinforcement learning rules, such as a variant of STDP called reward-modulated spike-timing-dependent plasticity (RSTDP). However, these previous models have assumed that no unsupervised learning is present (i.e., no learning occurs without, or independent of, rewards). We show that these models cannot elicit firing rate reinforcement while exhibiting both reward learning and ongoing, stable unsupervised learning. To address this issue, we propose a new RSTDP model of synaptic plasticity, based upon the observed effects that dopamine has on long-term potentiation and depression, that is able to exhibit unsupervised learning and lead to firing rate reinforcement.