Dandan Li

Yuan Zhou, Dandan Li, Shuwei Huo et al.

The choice of activation function in deep networks has a significant effect on the training dynamics and task performance. At present, the most effective and widely-used activation function is ReLU. However, because of the non-zero mean, negative missing and unbounded output, ReLU is at a potential disadvantage during optimization. To this end, we introduce a novel activation function to manage to overcome the above three challenges. The proposed nonlinearity, namely "Soft-Root-Sign" (SRS), is smooth, non-monotonic, and bounded. Notably, the bounded property of SRS distinguishes itself from most state-of-the-art activation functions. In contrast to ReLU, SRS can adaptively adjust the output by a pair of independent trainable parameters to capture negative information and provide zero-mean property, which leading not only to better generalization performance, but also to faster learning speed. It also avoids and rectifies the output distribution to be scattered in the non-negative real number space, making it more compatible with batch normalization (BN) and less sensitive to initialization. In experiments, we evaluated SRS on deep networks applied to a variety of tasks, including image classification, machine translation and generative modelling. Our SRS matches or exceeds models with ReLU and other state-of-the-art nonlinearities, showing that the proposed activation function is generalized and can achieve high performance across tasks. Ablation study further verified the compatibility with BN and self-adaptability for different initialization.

Dandan Li, Yuan Zhou, Shuwei Huo et al.

Convolutional neural networks (CNNs) are inherently suffering from massively redundant computation (FLOPs) due to the dense connection pattern between feature maps and convolution kernels. Recent research has investigated the sparse relationship between channels, however, they ignored the spatial relationship within a channel. In this paper, we present a novel convolutional operator, namely comb convolution, to exploit the intra-channel sparse relationship among neurons. The proposed convolutional operator eliminates nearly 50% of connections by inserting uniform mappings into standard convolutions and removing about half of spatial connections in convolutional layer. Notably, our work is orthogonal and complementary to existing methods that reduce channel-wise redundancy. Thus, it has great potential to further increase efficiency through integrating the comb convolution to existing architectures. Experimental results demonstrate that by simply replacing standard convolutions with comb convolutions on state-of-the-art CNN architectures (e.g., VGGNets, Xception and SE-Net), we can achieve 50% FLOPs reduction while still maintaining the accuracy.

Qijun Zhu, Dandan Li, Dik Lun Lee

As the web expands in data volume and in geographical distribution, centralized search methods become inefficient, leading to increasing interest in cooperative information retrieval, e.g., federated text retrieval (FTR). Different from existing centralized information retrieval (IR) methods, in which search is done on a logically centralized document collection, FTR is composed of a number of peers, each of which is a complete search engine by itself. To process a query, FTR requires firstly the identification of promising peers that host the relevant documents and secondly the retrieval of the most relevant documents from the selected peers. Most of the existing methods only apply traditional IR techniques that treat each text collection as a single large document and utilize term matching to rank the collections. In this paper, we formalize the problem and identify the properties of FTR, and analyze the feasibility of extending LSI with clustering to adapt to FTR, based on which a novel approach called Cluster-based Distributed Latent Semantic Indexing (C-DLSI) is proposed. C-DLSI distinguishes the topics of a peer with clustering, captures the local LSI spaces within the clusters, and consider the relations among these LSI spaces, thus providing more precise characterization of the peer. Accordingly, novel descriptors of the peers and a compatible local text retrieval are proposed. The experimental results show that C-DLSI outperforms existing methods.

Douglas Summers-Stay, Peter Sutor, Dandan Li

Representing meaning in the form of high dimensional vectors is a common and powerful tool in biologically inspired architectures. While the meaning of a set of concepts can be summarized by taking a (possibly weighted) sum of their associated vectors, this has generally been treated as a one-way operation. In this paper we show how a technique built to aid sparse vector decomposition allows in many cases the exact recovery of the inputs and weights to such a sum, allowing a single vector to represent an entire set of vectors from a dictionary. We characterize the number of vectors that can be recovered under various conditions, and explore several ways such a tool can be used for vector-based reasoning.