Files in this item

FilesDescriptionFormat

application/pdf

application/pdfJOHNSON-DISSERTATION-2016.pdf (7MB)
(no description provided)PDF

Description

Title:Minimum-error, energy-constrained source coding by sensory neurons
Author(s):Johnson, Erik C.
Director of Research:Jones, Douglas L.
Doctoral Committee Chair(s):Jones, Douglas L.
Doctoral Committee Member(s):Levinson, Stephen E.; Varshney, Lav R.; Nelson, Mark E.; Ratnam, Rama
Department / Program:Electrical & Computer Eng
Discipline:Electrical & Computer Engr
Degree Granting Institution:University of Illinois at Urbana-Champaign
Degree:Ph.D.
Genre:Dissertation
Subject(s):neural coding
sensory neurons
source coding
temporal coding
minimum energy
Abstract:Neural coding, the process by which neurons represent, transmit, and manipulate physical signals, is critical to the function of the nervous system. Despite years of study, neural coding is still not fully understood. Efforts to model neural coding could improve both the understanding of the nervous system and the design of artificial devices which interact with neurons. Sensory receptors and neurons transduce physical signals into a sequence of action potentials, called a spike train. The principles which underly the translation from signal to spike train are still under investigation. From the perspective of an organism, neural codes which maximize the fidelity of the encoded signal (minimize encoding error), provide a competitive advantage. Selective pressure over evolutionary timescales has likely encouraged neural codes which minimize encoding error. At the same time, neural coding is metabolically expensive, which suggests that selective pressure would also encourage neural codes which minimize energy. Based on these assumptions, this work proposes a principle of neural coding which captures the trade-off between error and energy as a constrained optimization problem of minimizing encoding error while satisfying a constraint on energy. A solution to the proposed optimization problem is derived in the limit of high spike-rates. The solution is to track the instantaneous reconstruction error, and to time spikes when the error crosses a threshold value. In the limit of large signals, the threshold level is a constant, but in general it is signal dependent. This coding model, called the neural source coder, implies neurons should be able to track reconstruction error internally, using the error signal to precisely time spikes. Mathematically, this model is similar to existing adaptive threshold models, but it provides a new way to understand coding by sensory neurons. Comparing the predictions of the neural source coder to experimental data recorded from a peripheral neuron, the coder is able to predict spike times with considerable accuracy. Intriguingly, this is also true for a cortical neuron which has a low spike-rate. Reconstructions using the neural source coder show lower error than other spiking neuron models. The neural source coder also predicts the asymmetric spike-rate adaptation seen in sensory neurons (the primary-like response). An alternative expression for the neural source coder is as an instantaneous-rate coder of a rate function which depends on the signal, signal derivative, and encoding parameters. The instantaneous rate closely predicts experimental peri-stimulus time histograms. The addition of a stochastic threshold to the neural source coder accounts for the spike-time jitter observed in experimental datasets. Jittered spike-trains from the neural source coder show long-term interval statistics which closely match experimental recordings from a peripheral neuron. Moreover, the spike trains have strongly anti-correlated intervals, a feature observed in experimental data. Interestingly, jittered spike-trains do not improve reconstruction error for an individual neuron, but reconstruction error is reduced in simulations of small populations of independent neurons. This suggests that jittered spike-trains provide a method for small populations of sensory neurons to improve encoding error. Finally, a sound coding method for applying the neural source coder to timing spikes for cochlear implants is proposed. For each channel of the cochlear implant, a neural source coder can be used to time pulses to follow the patterns expected by peripheral neurons. Simulations show reduced reconstruction error compared to standard approaches using the signal envelope. Initial experiments with normal-hearing subjects show that a vocoder simulating this cochlear implant sound coding approach results in better speech perception thresholds when compared to a standard noise vocoder. Although further experiments with cochlear implant users are critical, initial results encourage further study of the proposed sound-coding method. Overall, the proposed principle of minimum-error, energy-constrained encoding for sensory neural coding can be implemented by a spike-timing model with a feedback loop which computes reconstruction error. This model of neural source coding predicts a wide range of experimental observations from both peripheral and cortical neurons. The close agreement between experimental data and the predictions of the neural source coder suggests a fundamental principle underlying neural coding.
Issue Date:2016-07-08
Type:Thesis
URI:http://hdl.handle.net/2142/95741
Rights Information:Copyright 2016, Erik C. Johnson
Date Available in IDEALS:2017-03-16
Date Deposited:2016-08


This item appears in the following Collection(s)

Item Statistics