Learning Credit Assignment

参赛者：

李婵-中山大学

作品简介：

为了破解深度学习黑箱，建立神经网络决策与微观组分之间的桥梁，我们提出一种由稀疏度与均值方差刻画的随机连接权重模型，通过平均场方法训练得到系综水平上一组子网络，并在真实据集上获得了不弱于传统方法的表现。

Credit assignment problem (CAP) has long been an interesting topic connecting the macroscopic behaviors with microscopic interactions of components in deep neural networks. To solve this problem and reveal the mystery behind the black box of deep learning, we put forward a model with random weights characterized by a spike and slab distribution, which obtains comparable or even better performance than traditional models. An optimal random network ensemble can be achieved after training based on mean-field theory.