a08a33a57729d0f93c02ac9c963136d0eb704639
Recently, Lynch and Houghton (2015) have developed a method to predict in vivo response of auditory neurons using an estimated spectrotemporal receptive field (STRF) filter combined with an integrate-and-fire neuron model based on the Izhikevich neuron (Izhikevich, 2003). This model adds an additional adaptive parameter to the integrate-and-fire neuron for enhanced biophysical realism. Optimization is performed iteratively on the parameters of the filter and the parameters of the neuron model using a genetic algorithm in which fitness is evaluated by the van Rossum distance metric (van Rossum, 2001).
Lynch and Houghton’s method fills a clear need in the neuroinformatics field and moves forward the possibilities for neuron modeling in in vivo electrophysiology research. There exist many avenues for further improvement, including models with more biologically interpretable parameters and improved optimization algorithms.
In this paper, we propose a dynamical systems-based neuron model combined with an STRF filter that provides superior prediction accuracy with a computationally efficient optimization algorithm. This method is based on a Hindmarsh-Rose (HR) neuron model (Hindmarsh & Rose, 1984), which strikes a balance between the limited parameter sets of the integrate-and-fire models and the biological realism of the Hodgkin-Huxley ion current models (Hodgkin & Huxley, 1952). The feature space is explored by emcee (Foreman-Mackey et al., 2013), a Python implementation of a Markov chain Monte Carlo (MCMC), to find a local optima using the computational efficient SPIKE synchronization metric (Kreuz et al., 2015) as the fitness function. This combination of algorithms is first tested on simulated data with known parameters, and then validation on a real data set recorded in vivo from auditory neurons in the zebra finch (Taeniopygia guttata).
The HR-neuron model exhibits chaotic
characteristics that allow for multiple possible variations on spike times,
shape, and quantity. These chaotic dynamics are well studied for many possible parameter combinations (Storace,
Linaro, & de Lange, 2008; Shilnikov & Kolomeites, 2008).
Genetic
Algorithm
Mathematical
models such as the HR-neuron provide researchers a framework to understand and
predict qualities of a given system of interest. Researchers have developed
...
augmented HR-neuron model that includes a sensory filter (DSTRF).
We also estimated neuron and filter parameters using a Markov Chain Monte Carlo ("MCMC") technique. MCMC provides some distinct advantages over other parameter estimation methods, such as variational methods. First, MCMC provides an estimate for the full posterior distribution of the parameters rather than just a single value as with genetic algorithms. Having the posterior distribution for the parameters is useful for drawing inferences from the uncertainty of the parameter estimates. Secondly, MCMC allows for a Bayesian approach to estimating the parameters. Prior knowledge or beliefs about the parameters may be used in the estimation procedure and then updated afterwards. Finally, MCMC is simple to run in parallel on multi-core or multi-processor computers, which allows for significant reductions in run time.
To sample from the posterior distribution, we used emcee, a Python package implementing an affine-invariant ensemble sampler (Foreman-Mackey et al. 2013). The affine-invariant sampler is insensitive to covariances between parameters and requires tuning of much fewer hyper-parameters than standard MCMC algorithms (Goodman and Weare 2010). Further, the ensemble method used by the sampler was designed to run in a parallel processing environment. Rather than having one chain randomly sampling the posterior distribution, the ensemble sampler has hundreds of small chains sampling at once. These features should allow the sampler to converge quicker than standard MCMC algorithms on good parameter estimations
**Tyler - This is where we'll talk about SPIKy and stuff. I'm working on this part now.**