Decoupled Networks

By Weiyang Liu, Z. Liu, Zhiding Yu, Bo Dai, Rongmei Lin, Yisen Wang, James M. Rehg, Le Song. 2018

87 citations recorded by the source corpus.

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.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

Citations in the corpus

Continual Learning by Modeling Intra-Class Variation (2022, neutral): “Moreover, we empirically observe from Liu et al. (2016; 2018b); Wen et al. (2019) that the intra-class representations, when projected onto a unit hypersphere, are centered around the class mean and distributed like a vMF distribution.”
Angular Visual Hardness (2019, methodology): “Outcome: Failure to Reject Null Hypothesis [42] conjectures that ‖x‖2 accounts for intra-class Human/Model Confidence.”
Learning towards Minimum Hyperspherical Energy (2018, methodology): “. Therefore, applying weighted MHE to neural networks may reduce the recognition accuracy, further validating the importance of the hyperspherical (i.e., angular) diversity, as has been emphasized in [27, 29, 26]. 4 Theoretical Insights Different from a rigorous theoretical analysis, our goal here i”
SphereFace Revived: Unifying Hyperspherical Face Recognition (2021, background): “Technical Report, September 2021 tive on a hypersphere [8].”
Learning with Hyperspherical Uniformity (2021, background): “[51, 52, 49, 48, 50, 80, 79, 20, 19, 60, 14, 47] propose to learn representations on hypersphere and show that angular information in neural networks, in contrast to magnitude, preserves the key semantics and is very crucial to generalization.”
Kervolutional Neural Networks (2019, background): “ntity as activity vectors via a capsule structure. 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,”
Regular Polytope Networks (2021, background): “Several papers have followed this intuition and proposed to train DCNNs by direct angle optimization [23], [24], [25], [26].”
Regularizing Neural Networks via Minimizing Hyperspherical Energy (2019, methodology): “For all the experiments on CIFAR-10 and CIFAR-100 in the paper, we use the same data augmentation as [1, 34].”
Hyperspherical Prototype Networks (2019, unknown): “lasses. In similar fashion, several works project network outputs to the hypersphere for classiﬁcation through ‘ 2 normalization, which forces softmax cross-entropy to optimize for angular separation [12, 21, 22, 43, 44, 47]. Gidaris and Komodakis [10] show that using cosine similarity in the output”