Improving Neural ODE Training with Temporal Adaptive Batch Normalization

Authors: Su Zheng, Zhengqi Gao, Fan-Keng Sun, Duane Boning, Bei Yu, Martin D. Wong

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive numerical experiments on image classification and physical system modeling substantiate the superiority of TA-BN compared to baseline methods.
Researcher Affiliation Academia Su Zheng1 , Zhengqi Gao2 , Fan-Keng Sun2, Duane S. Boning2, Bei Yu1, Martin Wong1 1Department of CSE, CUHK 2 Department of EECS, MIT
Pseudocode Yes Algorithm 1 The forward pass of a TA-BN layer at time tj
Open Source Code Yes We put part of the code for reproduciblity in supplementary. It will be released upon acceptance.
Open Datasets Yes We conduct image classification across datasets including MNIST [26], SVHN [33], CIFAR-10, CIFAR-100 [22], and Tiny-Image Net [24].
Dataset Splits No The paper explicitly states a 90% training and remaining testing split for the Charge Pump circuit modeling dataset, but does not provide details for a validation split for any of the datasets used.
Hardware Specification Yes All experiments are run on a Linux server with RTX 3090 GPUs.
Software Dependencies No The paper mentions software like 'Py Torch' and 'Torch Diffeq' but does not specify their version numbers.
Experiment Setup Yes We employ the dopri5 solver with a tolerance of 10 3 for ODE solving and adopt the Adam W optimizer [27] with a learning rate of 10 3 to train the neural networks for 128 epochs. The training batch size is 256. We set M = 100 for TA-BN.