Neural Kernels Without Tangents

Authors: Vaishaal Shankar, Alex Fang, Wenshuo Guo, Sara Fridovich-Keil, Jonathan Ragan-Kelley, Ludwig Schmidt, Benjamin Recht

ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimentally, we show a correlation in test error between neural network architectures and the associated kernels. We construct a simple neural network architecture using only 3 3 convolutions, 2 2 average pooling, Re LU, and optimized with SGD and MSE loss that achieves 96% accuracy on CIFAR10, and whose corresponding compositional kernel achieves 90% accuracy.
Researcher Affiliation Academia 1University of California, Berkeley 2Massachusetts Institute of Technology.
Pseudocode Yes Algorithm 1 Compositional Kernel
Open Source Code No The paper does not include an unambiguous statement that the authors are releasing their code for the described methodology, nor does it provide a direct link to a source code repository.
Open Datasets Yes We then present comparison results between neural networks, NTKs, and compositional kernels on a variety of datasets, including MNIST, CIFAR-10 (Krizhevsky (2009)), CIFAR-10.1 (Recht et al. (2019)), CIFAR-100 (Krizhevsky (2009)) and 90 UCI datasets (Fern andez-Delgado et al. (2014)).
Dataset Splits Yes Table 3 compares the performance of neural networks with various depths and their corresponding compositional kernels on both the 10,000 test images from CIFAR-10 and the additional 2,000 harder test images from CIFAR-10.1. We compute the optimal hyperparameters for each dataset (for both NTK and Gaussian kernel) by averaging performance over four cross-validation folds.
Hardware Specification Yes We implemented all the convolutional kernels in the tensor comprehensions framework (Vasilache et al., 2018) and executed them on V100 GPUs using Amazon Web Services (AWS) P3.16xlarge instances.
Software Dependencies No The paper mentions implementing kernels in the 'tensor comprehensions framework' and refers to a publication (Vasilache et al., 2018), but it does not provide specific version numbers for this framework or any other software dependencies crucial for replication.
Experiment Setup Yes We train all the Myrtle CNNs on CIFAR-10 using SGD and the mean squared error (MSE) loss with multi-step learning rate decay. The exact hyperparameters are provided in the appendix.