Towards Accurate BNNs via Modeling Contextual Dependencies

The proposed binary model is built on modeling contextual dependencies. An overview of the BCDNet architecture is illustrate in Fig. 4. We first describe the fundamentals of binary MLP block. Then, we introduce the contextual design of binary convolutions with dynamic embeddings. Finally, we present the complete BCDNet architecture with modeling contextual dependencies.

本研究提出的二值模型基于上下文依赖建模构建而成。BCDNet架构的整体框架如图4所示。本文首先阐述二值多层感知机（Multi-Layer Perceptron，简称MLP）模块的基本原理，随后介绍结合动态嵌入的二值卷积的上下文设计方案，最后完整呈现具备上下文依赖建模能力的BCDNet架构。