Fitting of dynamic recurrent neural network models to sensory stimulus-response data

Area of Science:

• Computational Neuroscience
• Machine Learning for Neuroscience

Background:

• Conventional neural network training is unsuitable for sensory neuron data due to the discrete nature of spike timings.
• Sensory neuron responses lack amplitude information, posing challenges for direct data fitting.

Purpose of the Study:

• To develop and validate a recurrent neural network (RNN) model for fitting sensory neuron stimulus-response data.
• To address the challenge of non-continuous neural spike timing data in model fitting.

Main Methods:

• Utilized a recurrent dynamical neuron network model with universal approximation properties.
• Employed maximum likelihood estimation (MLE) with a likelihood function derived from Poisson statistics of neural spiking.
• Generated stimulus data using a phased cosine Fourier series with varying amplitude, component size, and sample size.

Main Results:

• The RNN model successfully fitted sensory neuron data, demonstrating its capability to capture excitatory-inhibitory characteristics.
• The study analyzed the impact of stimulus parameters (amplitude, component size, sample size) on the model identification process.
• Model performance was validated through comparisons with studies using identical and different modeling approaches.

Conclusions:

• The proposed RNN model and MLE approach provide a robust framework for fitting sensory neuron data.
• The findings highlight the effectiveness of this method in characterizing complex neural dynamics from spike timing information.
• This theoretical study offers a valuable tool for advancing our understanding of sensory neuron computation.