Constrained Optimization to Train Neural Networks on Critical and Under-Represented Classes

Authors: Sara Sangalli, Ertunc Erdil, Andeas Hötker, Olivio Donati, Ender Konukoglu

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

Reproducibility Variable Result LLM Response
Research Type Experimental We present experimental results for image-based binary and multi-class classification applications using an in-house medical imaging dataset, CIFAR10, and CIFAR100.
Researcher Affiliation Academia 1 Computer Vision Lab, ETH Zürich 2 Institute for Diagnostic and Interventional Radiology, Universitätsspital Zürich
Pseudocode Yes Algorithm 1 ALM for Training DNNs
Open Source Code Yes 1Code is available at: https://github.com/salusanga/alm-dnn.
Open Datasets Yes CIFAR10 and CIFAR100 [11]
Dataset Splits Yes In our experiments, we randomly split 20% of the training cohort as validation set by keeping the class imbalance consistent across the datasets.
Hardware Specification Yes The proposed method is implemented in Py Torch and we run all experiments on a Nvidia Ge Force GTX Titan X GPU with 12GB memory.
Software Dependencies No The paper states 'The proposed method is implemented in Py Torch' but does not specify the version number for PyTorch or any other software dependencies.
Experiment Setup Yes Hyper-parameters selection: Selection of the best hyper-parameters is crucial both to ensure proper and fair evaluation of the methods and to understand the true performance of any model. To achieve this, we perform grid-search to determine the hyper-parameters that yield the highest AUC for the binary experiments. ... In the proposed method, there are 4 parameters to be set: µ(0), λ(0), ρ, and δ. Thus, hyperparameters search is an important aspect of the proposed method. ... We initialize all the Lagrangian multipliers λ(0) i to 0. We choose µ(0) from the set {10 7, 10 6, 10 5, 10 4, 10 3}, as it is suggested to choose a small value in the beginning and increase it iteratively using the equation µ(k+1) = ρ µ(k). We choose ρ from the set {2, 3} as ρ > 1 is suggested.