Topologically Densified Distributions

Authors: Christoph Hofer, Florian Graf, Marc Niethammer, Roland Kwitt

ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental we provide empirical evidence (across various vision benchmarks) to support our claim for better generalization. For our experiments, we draw on a setup common to many works in semi-supervised learning (...) In particular, we present experiments on three (10 class) vision benchmark datasets: MNIST, SVHN and CIFAR10.
Researcher Affiliation Academia 1Department of Computer Science, Univ. of Salzburg, Austria 2Univ. of North Carolina, Chapel Hill, USA.
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes Py Torch source code is available at https://github.com/c-hofer/topologically_densified_distributions
Open Datasets Yes In particular, we present experiments on three (10 class) vision benchmark datasets: MNIST, SVHN and CIFAR10.
Dataset Splits Yes Thus, we study the behavior of crossvalidating β, when the validation set is of size equal to the training corpus. The shaded region shows the variation in the testing error on small-validation sets.
Hardware Specification No The paper does not specify any particular hardware (e.g., GPU, CPU models, or memory) used for running the experiments.
Software Dependencies No The paper mentions 'Py Torch source code' but does not specify a version number or list other software dependencies with their versions.
Experiment Setup Yes Optimization is done by SGD with momentum (0.9) over 310 epochs with cross-entropy loss and cosine learning rate annealing (Loshchilov & Hutter, 2017) (without restarts). As all experiments use weight decay, it is important to note that batch normalization combined with weight decay mainly affects the effective learning rate (van Laarhoven, 2017; Zhang et al., 2019). (...) We choose a sub-batch size of b = 16 and draw n = 8 sub-batches (see 2.4); this amounts to a total batch size of 128.