Utilizing Class Information for Deep Network Representation Shaping

Authors: Daeyoung Choi, Wonjong Rhee3396-3403

AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We performed an extensive set of experiments on the most popular datasets (MNIST, CIFAR-10, CIFAR-100) and architectures (MLP, CNN). Additionally, Res Net (He et al. 2016) was tested as an example of a sophisticated network, and an image reconstruction task using autoencoder was tested as an example of a different type of task. We have tested a variety of scenarios with different optimizers, number of classes, network size, and data size. The results show that our representation regularizers outperform the baseline (no regularizer) and L1/L2 weight regularizers for almost all the scenarios that we have tested.
Researcher Affiliation Academia Daeyoung Choi, Wonjong Rhee Department of Transdisciplinary Studies Seoul National University Seoul, 08826, South Korea {choid, wrhee}@snu.ac.kr
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes Code is made available at https://github.com/snu-adsl/classwiseregularizer.
Open Datasets Yes Three popular datasets (MNIST, CIFAR-10, and CIFAR100) were used as benchmarks.
Dataset Splits Yes When a regularizer (including L1W and L2W) was used for an evaluation scenario, the penalty loss weight λ was determined as one of {0.001, 0.01, 0.1, 1, 10, 100} using 10,000 validation samples.
Hardware Specification No The paper does not provide specific details about the hardware (e.g., GPU, CPU models) used for running the experiments.
Software Dependencies No The paper does not list specific software dependencies with version numbers.
Experiment Setup Yes When a regularizer (including L1W and L2W) was used for an evaluation scenario, the penalty loss weight λ was determined as one of {0.001, 0.01, 0.1, 1, 10, 100} using 10,000 validation samples... Mini-batch size was increased to 500 for CIFAR-100 such that classwise operations can be appropriately performed but was kept at the default value of 100 for MNIST and CIFAR-10.