Dissecting Deep Learning Networks—Visualizing Mutual Information

Deep Learning (DL) networks are recent revolutionary developments in artificial intelligence research. Typical networks are stacked by groups of layers that are further composed of many convolutional kernels or neurons. In network design, many hyper-parameters need to be defined heuristically before training in order to achieve high cross-validation accuracies. However, accuracy evaluation from the output layer alone is not sufficient to specify the roles of the hidden units in associated networks. This results in a significant knowledge gap between DL’s wider applications and its limited theoretical understanding. To narrow the knowledge gap, our study explores visualization techniques to illustrate the mutual information (MI) in DL networks. The MI is a theoretical measurement, reflecting the relationship between two sets of random variables even if their relationship is highly non-linear and hidden in high-dimensional data. Our study aims to understand the roles of DL units in classification performance of the networks. Via a series of experiments using several popular DL networks, it shows that the visualization of MI and its change patterns between the input/output with the hidden layers and basic units can facilitate a better understanding of these DL units’ roles. Our investigation on network convergence suggests a more objective manner to potentially evaluate DL networks. Furthermore, the visualization provides a useful tool to gain insights into the network performance, and thus to potentially facilitate the design of better network architectures by identifying redundancy and less-effective network units