Deep Combinatorial Aggregation

Authors: Yuesong Shen, Daniel Cremers

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

Reproducibility Variable Result LLM Response
Research Type Experimental We experiment on in-domain, distributional shift, and out-of-distribution image classification tasks, and empirically confirm the effectiveness of DCWA and DCA approaches.
Researcher Affiliation Academia Yuesong Shen 1,2 Daniel Cremers 1,2 1 Technical University of Munich, Germany 2 Munich Center for Machine Learning, Germany {yuesong.shen, cremers}@tum.de
Pseudocode Yes In Appendix A we provide a pseudo-code for layerwise DCA training.
Open Source Code Yes Source code is available at https://github.com/tum-vision/dca.
Open Datasets Yes For this we use CIFAR-10 [22] and SVHN [31] datasets
Dataset Splits No The paper uses standard datasets like CIFAR-10 and SVHN but does not explicitly provide the specific percentages or counts for training, validation, and test splits.
Hardware Specification Yes Our implementation uses Py Torch [32] and can run on a single modern GPU with 10Gb VRAM.
Software Dependencies No Our implementation uses Py Torch [32] - while PyTorch is mentioned, a specific version number is not provided, which is required for reproducible software dependencies.
Experiment Setup Yes For all training we use SGD with momentum 0.9. We use a drop rate of 0.1 for MC dropout, and for SWAG we follow the settings from Maddox et al. [28]. Deep ensemble is performed on five separate base networks. And trunkwise and modelwise DCA models also use five copies. We schedule 200 epochs for standard baseline training. Then, for training a DCA model with three sets of component instances we use 600 epochs, and accumulate for each minibatch an average gradient over three backpropagations on randomly selected DCA proposals for parameter update.