Decoupled Networks

Abstract

Inner product-based convolution has been a central component of convolutional neural networks (CNNs) and the key to learning visual representations. Inspired by the observation that CNN-learned features are naturally decoupled with the norm of features corresponding to the intra-class variation and the angle corresponding to the semantic difference, we propose a generic decoupled learning framework which models the intra-class variation and semantic difference independently. Specifically, we first reparametrize the inner product to a decoupled form and then generalize it to the decoupled convolution operator which serves as the building block of our decoupled networks. We present several effective instances of the decoupled convolution operator. Each decoupled operator is well motivated and has an intuitive geometric interpretation. Based on these decoupled operators, we further propose to directly learn the operator from data. Extensive experiments show that such decoupled reparameterization renders significant performance gain with easier convergence and stronger robustness.

Kervolutional Neural Networks

“…This opens a new research space for artificial neural networks, although its performance on large dataset is still relatively weak. Decoupled network [38] interprets convolution as the product of norm and cosine angle of the weight and input vectors, resulting in explicit geometrical modeling of the intra-class and extra-class variation. To process graph inputs, Spline-CNN [14] extends convolution by using continuous B-spline basis, which is parametrized by a constant number of trainable control values.…”

Learning Robust Embedding Representation With Hybrid Loss for Classification and Verification

“…Liu et al [28] concluded them as a deep hyperspherical learning framework which forces the norms to 1 in order to make the outputs reflected the angles. And Liu et al [29] prove the idea further that not all norms are equal to 1 via decoupling the convolutional operator. In such a way, different kinds of functions rely on norm and angle are considered to substitute the functions of magnitude and angular.…”

“…Even though the tradeoff between invariance and discriminativeness is unknown, we still have chances to balance the two properties above in the manner of deep metric learning. According to the recent work [22], [25], [28], [29], inter-class variation and intra-class variation can be decoupled by relating to the norm and the angle within an inner product operation respectively. They argue that the angle corresponds to the inter-class variation and the norm accounts for intra-class variation.…”

Fast and Accurate Lane Detection via Graph Structure and Disentangled Representation Learning

“…For the third difficulty, to pursue higher running speed, we also drop the decoder network and compress the high-dimensional feature map as above works. However, unlike they directly use 1 × 1 convolution to compress features, we utilize the idea of disentangled representation learning [ 21 , 22 ] to restain more information in compressed features and name this module as Disentangled Feature Compressor (DFC). Specifically, DFC divides the high-dimensional feature map into three groups and integrates features respectively through low-dimensional 1 × 1 convolution.…”