Linear-nonlinear-Poisson (LPN) model provides a good interpretation of neuron response

and one of its important links is linear feature extraction. Focusing on the issue that traditional information-theoretic Spike-Triggered average and covariance (iSTAC) algorithm cannot exactly describe neuron feature and extract its motion characteristics

especially in low-dimension stimulus

this study improves the iSTAC algorithm and proposes a new algorithm for neuron-filtering feature extraction. Statistics of non-spike-triggered stimulus are introduced to build a highly accurate objective function of neuron-filtering feature subspace and to enhance noise resistance of the system. To optimize solution space and to accelerate convergence rate

a variable metric method is accepted to maximize the object function. Experimental results of linear filter recovery under different nonlinear conditions show that the new algorithm is similar to the traditional iSTAC algorithm in high-dimension stimulus and has achieved significant progress in less than 6 500 dimension stimuli. Furthermore

results show that the new algorithm is better than the spike-triggered average (STA) and spike-triggered covariance (STC) algorithms. The proposed new algorithm has better adaptability and greater robustness

and can be applied to establish a complete extraction model of video motion feature based on visual features.