Awesome

On the Equivalence between Neural Network and Support Vector Machine

Codes for NeurIPS 2021 paper "On the Equivalence between Neural Network and Support Vector Machine".

Overview

In this paper, we prove the equivalence between neural network (NN) and support vector machine (SVM), specifically, the infinitely wide NN trained by soft margin loss and the standard soft margin SVM with NTK trained by subgradient descent. Our main theoretical results include establishing the equivalence between NN and a broad family of L2 regularized kernel machines (KMs) with finite-width bounds, which cannot be handled by prior work, and showing that every finite-width NN trained by such regularized loss functions is approximately a KM.

Furthermore, we demonstrate our theory can enable three practical applications, including

• non-vacuous generalization bound of NN via the corresponding KM;
• non-trivial robustness certificate for the infinite-width NN (while existing robustness verification methods (e.g. IBP, Fast-Lin, CROWN) would provide vacuous bounds);
• intrinsically more robust infinite-width NNs than those from previous kernel regression.

See our paper and slides for details.

Code overview

• train_sgd.py: train the NN and SVM with NTK with stochastic subgradient descent. Plot the results to verify the equivalence.

• regression.py: kernel ridge regression with NTK.

• robust_svm.py:

  • test(): evaluate the robustness of NN using IBP or SVM with our method in the paper.
  • test_regressions(): evaluate the robustness of kernel ridge regression models using our method.
  • bound_ntk()：calculate the lower and upper bound for NTK of two-layer fully-connected NN.

• ibp.py: functions to calculate IBP bounds. Specified for NTK parameterization.

• models/model.py: codes for constructing fully-connected neural networks with NTK parameterization.

• config/:

  • svm_sgd.yaml: configurations and hyper-parameters to train NN and SVM.
  • svm_gene.yaml: configurations and hyper-parameters to calculate generalization bound.

Required environments:

This code is tested on the below environments:

python==3.8.8
torch==1.8.1
neural-tangents==0.3.6

Other required packages can be installed using Conda as follows,

conda create -n equiv-nn-svm python=3.8
conda activate equiv-nn-svm
conda install numpy tqdm matplotlib seaborn pyyaml

For the installation of PyTorch, please reference the instructions from https://pytorch.org/get-started/locally/. For the installation and usage of neural-tangents, please reference the instructions at https://github.com/google/neural-tangents.

Experiments

Train NN and SVM to verify the equivalence

python train_sgd.py

Example of the SGD results

Example of the GD results

Robustness verification of NN

Add your paths to your NN models in the code and separate by the width. Specify the width of the models you want to verify. Then run the test() function in robust_svm.py.

python -c "import robust_svm; robust_svm.test('nn')"

Robustness verification of SVM

Add your paths to your SVM models in the code. Then run the test() function in robust_svm.py.

python -c "import robust_svm; robust_svm.test('svm')"

Train kernel ridge regression with NTK models

python regression.py

Robustness verification of kernel ridge regression models

Run test_regressions() function in robust_svm.py.

python -c "import robust_svm; robust_svm.test_regressions()"

Cite our paper

Yilan Chen, Wei Huang, Lam M. Nguyen, Tsui-Wei Weng, "On the Equivalence between Neural Network and Support Vector Machine", NeurIPS 2021.

@inproceedings{chen2021equivalence,
  title={On the Equivalence between Neural Network and Support Vector Machine},
  author={Chen, Yilan and Huang, Wei and Nguyen, Lam M and Weng, Tsui-Wei},
  booktitle={Thirty-Fifth Conference on Neural Information Processing Systems},
  year={2021}
}