树突神经网络分析

Analysis on dendritic neural network

树突神经网络(DENN)是一种含有局部非线性树突结构的特殊神经网络模型。文章研究了DENNs在网络结构变化的情况下模型的学习行为。在有监督学习任务的实验中发现,局部非线性结构的DENNs可以提高模型的表达能力,并且在中等树突分支数量时表达能力最强,在网络较小的情况下DENNs模型比常规前馈型神经网络的优势表现得更加明显。在随机噪声数据集上的实验中发现DENNs拟合能力的优势不明显,这种现象进一步表明,DENNs的容量优势与自然图像数据中的冗余有关。

Dendritic Neural Network(DENN) is a special kind of network with localized nonlinearity. This work studies the learning behavior of DENNs when the network architecture is altered on various aspects. In the experiments of supervised machine learning task, it is found that the locality structure of DENNs can improve the expressive ability, and the expressive ability is strongest when DENNs with mid-size dendrite branch numbers, and DENNs show even greater advantage over the standard feedforward neural networks when network sizes are small. It is found that the fitting ability of DENNs is not obvious in the experiment on noise data learning task, this phenomenon further indicates that the improved model expressivity of DENNs owe to the structural redundancy of natural image data.