Real-time classification through a spiking deep belief network with intrinsic plasticity

Abstract

Deep Belief Networks (DBNs) has made a good effect in machine learning and object classification. However, the current question is how to reduce the computational cost without detrimental to accuracy. To solve this problem, this paper is undertaken to convert the Siegert neuron into LIF neuron in DBNs and analyze the effects of changing the value of parameters for spiking neurons such as thresholds and firing rates. Besides, we also add intrinsic plasticity (IP) into the network to render better adaptive capability. Besides, the most exciting results is the spiking DBN with intrinsic plasticity submits its first correct guess of the output label within an average of 2.5 ms after the onset of the simulated Poisson spike train input with the initial firing rates beyond 200 Hz, and the recognition accuracy is still more than 94 percent.