Learning with Retrospection

Authors: Xiang Deng, Zhongfei Zhang7201-7209

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on several benchmark datasets demonstrate the superiority of LWR for training DNNs.
Researcher Affiliation Academia Xiang Deng, Zhongfei Zhang Computer Science Department, State University of New York at Binghamton xdeng7@binghamton.edu, zhongfei@cs.binghamton.edu
Pseudocode Yes Algorithm 1 LWR
Open Source Code No The paper mentions using 'the open-source implementation 2 of (Zhang et al. 2017)' for robustness experiments, but it does not provide concrete access to the source code for the proposed Learning with Retrospection (LWR) method, nor does it state that its code is publicly available.
Open Datasets Yes We report the results on several benchmark datasets, i.e., CIFAR-10 (Krizhevsky and Hinton 2009), CIFAR-100 (Krizhevsky and Hinton 2009), Tiny Image Net 1, CUB-200-2011 (Wah et al. 2011), Stanford Dogs (Khosla et al. 2011), FGVC-Aircraft (Maji et al. 2013), Abalone (Dua and Graff 2017), Arcene (Dua and Graff 2017), and Iris (Dua and Graff 2017).
Dataset Splits No The paper states, 'CIFAR-10 is a 10-class image classification dataset, containing 50,000 training images and 10,000 test images,' and 'We follow the default training and test splits' for tabular datasets. While it mentions 'training and validation accuracies' generally, it does not provide specific details (percentages, counts, or explicit methodology) for a validation dataset split needed for reproduction across all experiments.
Hardware Specification No The paper does not provide specific details regarding the hardware used for running its experiments, such as GPU/CPU models, processor types, or memory amounts.
Software Dependencies No The paper mentions optimizers like 'SGD with momentum 0.9 and weight decay 5e-4' and 'Adam (Kingma and Ba 2015)', but it does not specify version numbers for any programming languages, libraries, or other key software components used in the experiments.
Experiment Setup Yes Following the standard training procedure for modern DNNs, we have trained all the networks for 200 epochs with optimizer SGD with momentum 0.9 and weight decay 5e-4 on CIFAR, CUB-200-2011, Stanford Dogs, and FGVC-Aircraft, 120 epochs for Tiny Image Net. We train it for 50 epochs with mini-batche size 16 by using Adam (Kingma and Ba 2015) with default hyperparameters.