DeepVisualInsight: Time-Travelling Visualization for Spatio-Temporal Causality of Deep Classification Training

Authors: Xianglin Yang, Yun Lin, Ruofan Liu, Zhenfeng He, Chao Wang, Jin Song Dong, Hong Mei5359-5366

AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our extensive experiments show that, comparing to baseline approaches, we achieve the best visualization performance regarding the spatial/temporal properties and visualization efficiency. Moreover, our case study shows that our visual solution can well reflect the characteristics of various training scenarios, showing good potential of DVI as a debugging tool for analyzing deep learning training processes.
Researcher Affiliation Collaboration 1School of Computing, National University of Singapore, Singapore 2Key Lab of High-Confidence Software Technology, Mo E (Peking University), China
Pseudocode No The paper describes algorithms and methods but does not present them in a structured pseudocode block or a clearly labeled algorithm section.
Open Source Code Yes More details of our tool/experiments are at (DVI 2021). DVI. 2021. DVI (Anonymous). https://sites.google.com/ view/deepvisualinsight/. Accessed: 2021-03-17.
Open Datasets Yes We choose three datasets, i.e., MNIST, Fashion-MNIST, and CIFAR-10. We use Res Net18 (He et al. 2016) as the subject classifier
Dataset Splits No The paper mentions training, testing, but does not explicitly state how validation data splits are used or if a separate validation set is employed for hyperparameter tuning. It refers to "subset training/testing sizes are set to 1000/200" for Deep View comparisons, but no general explicit split.
Hardware Specification No The paper does not specify the hardware used for running experiments (e.g., CPU, GPU models, memory).
Software Dependencies No The paper mentions using ResNet18 and common deep learning concepts but does not specify any software libraries or frameworks with their version numbers (e.g., PyTorch, TensorFlow, scikit-learn versions).
Experiment Setup Yes Given the dimension of the feature vector is h, we let the encoder and decoder to have shape (h, h / 2 , 2); and (2, h / 2 , h) respectively. Learning rate is initialized with 0.01 and decay every 8 epochs by factor of 10. The threshold δ to decide boundary point is set to be 0.1. We generate 0.1 N boundary points, shared by all the solutions. The upper bound for λ in boundary point generation is set to 0.4, α in Equation 5 to 0.8, β in Equation 8 to 1.0, and the trade-off hyper-parameters in total loss (Equation 12) to 1.0, 1.0, 0.3 respectively.