A physics perspective to understand credit assignment

Deep learning has been playing a seminal role in a variety of fields, while the Achilles heel of it lies in the black-box property hidden behind it, i.e., it is still unknown how the learning coordinates a huge number of parameters to achieve decision making. To explain hierarchical credit assignment, we propose a mean-field learning model by assuming that an ensemble of sub-networks, rather than a single network, is trained in a supervised scenario. The model works well and we identify three kinds of weights: VIP(important), UIP(unimportant), and other connections allowing for a broad distribution of their values, which provides insights toward understanding the macroscopic behavior of deep learning through distinct roles of synaptic weights. Then, we extend this model to learn the harder temporal credit assignment framed in the recurrent neural networks, which are widely used in processing complex temporal sequences. We successfully reveal important connections determining the overall performance of the network and produce an ensemble of candidate networks. Our method links network statistics, distinct functions of computational layers and neural selectivity, and can be used as a general and promising tool to understand credit assignment in networks with various architectures.