In this study, we aim to interpret the hidden semantics of each unit within convolutional neural networks by abstracting local activated patterns of each neuron into corresponding global functions. Unlike existing quantitative hidden-semantics-based explanations which require comparing pixel-wise annotated data one by one, the proposed active interpretability method gives function embeddings after the training without semantic annotation. Specifically, the function of a neuron is denoted as its global expected activated pattern, and therefore feature space and function embedding space are unsupervised aligned during training. The synchronization mechanism is introduced to aggregate scattered function embeddings into function modules, transforming the excessive gray-box interpretations into white-box ones. Moreover, the hard routing guided by function embedding is employed to ensure semantic specificity. We explore the aggregated function modules to showcase the qualitative interpretability of functionally motivated networks. Meanwhile, the proposed method exhibits superior quantitative interpretability metrics such as accuracy, faithfulness, robustness, and complexity.

Synchronization is a ubiquitous phenomenon in nature that enables the orderly presentation of information. In the human brain, for instance, functional modules such as the visual, motor, and language cortices form through neuronal synchronization. Inspired by biological brains and previous neuroscience studies, we propose an interpretable neural network incorporating a synchronization mechanism. The basic idea is to constrain each neuron, such as a convolution filter, to capture a single semantic pattern while synchronizing similar neurons to facilitate the formation of interpretable functional modules. Specifically, we regularize the activation map of a neuron to surround its focus position of the activated pattern in a sample. Moreover, neurons locally interact with each other, and similar ones are synchronized together during the training phase adaptively. Such local aggregation preserves the globally distributed representation nature of the neural network model, enabling a reasonably interpretable representation. To analyze the neuron interpretability comprehensively, we introduce a series of novel evaluation metrics from multiple aspects. Qualitative and quantitative experiments demonstrate that the proposed method outperforms many state-of-the-art algorithms in terms of interpretability. The resulting synchronized functional modules show module consistency across data and semantic specificity within modules.

Distributed data stream mining has gained increasing attention in recent years since many organizations collect tremendous amounts of streaming data from different locations. Existing studies mainly focus on learning evolving concepts on distributed data streams, while the privacy issue is little investigated. In this article, for the first time, we develop a federated learning framework for distributed concept-drifting data streams, called FedStream. The proposed method allows capturing the evolving concepts by dynamically maintaining a set of prototypes with error-driven representative learning. Meanwhile, a new metric-learning-based prototype transformation technique is introduced to preserve privacy among participating clients in the distributed data streams setting. Extensive experiments on both real-world and synthetic datasets have demonstrated the superiority of FedStream, and it even achieves competitive performance with state-of-the-art distributed learning methods.

Multi-label classification problem has gained growing attention in recent years due to its diverse applications to real-world problems such as image annotation and query suggestions. However, traditional multi-label classification methods tend to fail due to the inconsistency between input and output space, where similar instances in the feature space may have distinct semantic labels in the output space. To eliminate the inconsistency problem, in this paper, we propose a supervised metric learning approach for multi-label classification, called MLMLI, which attempts to learn a similarity metric for multi-label data. The basic idea is to incorporate label similarity in output space as weak supervision to assign higher similarity to the pairs of instances with more similar labels. To this end, a weighted triple loss, and a step-specified coordinate descent method are employed. Different from traditional dimensionality reduction approaches, MLMLI is independent of any prior information of data, and thus enjoys a high capacity of generalization. Moreover, the metric learned by MLMLI offers a new venue for feature learning. Experiments on real-world datasets have further demonstrated the effectiveness of MLMLI and show its superiority over many state-of-the-art algorithms.