SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2305.13948 · NeurIPS · 2024

Decoupled Kullback-Leibler Divergence Loss

Hanwang Zhang, Xiaojuan Qi, Zhuotao Tian, Jiequan Cui, Zhisheng Zhong, Bei Yu

arXiv · PDF · Open in the Atlas

Code that ran

We lifted 11 functions out of this paper's own repositories and ran 7 of them in a sandbox. "Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results.

RepositoryRoleRan
jiequancui/dkl canonical 3 of 3
jiequancui/LBGAT canonical 0 of 4
FPNAS/LBGAT pwc_unofficial 2 of 2
jiequancui/DKL — 1 of 1
dvlab-research/LBGAT — 1 of 1
FunctionStatusWhere it lives
IKL_KD Ran jiequancui/DKL/DKLv1/KD-dkl/mdistiller/distillers/IKL_KD.py
pointer only (licence: NONE) · get_code("b32e10adc9bf6201")
cross_entropy Ran jiequancui/dkl/DKLv1/Adv-training-dkl/train_dkl_cifar10.py
pointer only (licence: NONE) · get_code("59f6b4ce1de254e2")
dkl_loss Ran jiequancui/dkl/DKLv1/Adv-training-dkl/train_dkl_cifar10.py
pointer only (licence: NONE) · get_code("eb1d0d5bb0f0c627")
dkl_loss Ran jiequancui/dkl/DKLv1/KD-dkl/mdistiller/distillers/IKL_KD.py
pointer only (licence: NONE) · get_code("6ae3a3d2a8a1acc4")
l2_norm Ran FPNAS/LBGAT/lbgat.py
code served (permissive licence) · get_code("bbffec6aadbbfd95")
lbgat_loss Ran dvlab-research/LBGAT/lbgat.py
code served (permissive licence) · get_code("342a07cb13a5e0bc")
squared_l2_norm Ran FPNAS/LBGAT/lbgat.py
code served (permissive licence) · get_code("724c88e3f81d2123")
ResNet152_cifar Not yet run jiequancui/LBGAT/models/resnetcifar.py
code served (permissive licence) · get_code("87fe041ab6db516b")
ResNet18_cifar Not yet run jiequancui/LBGAT/models/resnetcifar.py
code served (permissive licence) · get_code("60d3d2bb8b6c30e2")
ResNet34_cifar Not yet run jiequancui/LBGAT/models/resnetcifar.py
code served (permissive licence) · get_code("4386c6d158344286")
lbgat_loss Not yet run jiequancui/LBGAT/lbgat.py
code served (permissive licence) · get_code("2533491b22247ed9")

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

In this paper, we delve deeper into the Kullback-Leibler (KL) Divergence loss and mathematically prove that it is equivalent to the Decoupled Kullback-Leibler (DKL) Divergence loss that consists of 1) a weighted Mean Square Error (wMSE) loss and 2) a Cross-Entropy loss incorporating soft labels. Thanks to the decomposed formulation of DKL loss, we have identified two areas for improvement. Firstly, we address the limitation of KL/DKL in scenarios like knowledge distillation by breaking its asymmetric optimization property. This modification ensures that the wMSE component is always effective during training, providing extra constructive cues. Secondly, we introduce class-wise global information into KL/DKL to mitigate bias from individual samples. With these two enhancements, we derive the Improved Kullback-Leibler (IKL) Divergence loss and evaluate its effectiveness by conducting experiments on CIFAR-10/100 and ImageNet datasets, focusing on adversarial training, and knowledge distillation tasks. The proposed approach achieves new state-of-the-art adversarial robustness on the public leaderboard -RobustBench and competitive performance on knowledge distillation, demonstrating the substantial practical merits. Our code is available at https://github.com/jiequancui/DKL.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2305.13948")
get_code_for_paper("2305.13948")
have("2305.13948")

Connect an agent — have() is free.