Poisson Learning: Graph Based Semi-Supervised Learning At Very Low Label Rates

Authors: Jeff Calder, Brendan Cook, Matthew Thorpe, Dejan Slepcev

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

Reproducibility Variable Result LLM Response
Research Type Experimental We present numerical experiments with semi-supervised classification on the MNIST, Fashion MNIST, and Cifar-10 datasets.
Researcher Affiliation Academia 1School of Mathematics, University of Minnesota, Minneapolis, USA. 2Department of Mathematics, University of Manchester, Manchester, UK. 3Department of Mathematical Sciences, Carnegie Mellon University, Pittsburgh, USA.
Pseudocode Yes Algorithm 1 Poisson Learning, Algorithm 2 Poisson MBO
Open Source Code Yes Source Code: https://github.com/jwcalder/Graph Learning
Open Datasets Yes We tested Poisson learning on three datasets: MNIST (Le Cun et al., 1998), Fashion MNIST (Xiao et al., 2017) and Cifar-10 (Krizhevsky et al., 2009).
Dataset Splits No No explicit details on validation set splits (percentages or counts) or specific methodology for creating such a split were found. The paper mentions random trials for labeled data but not a defined validation split.
Hardware Specification Yes We used an NVIDIA RTX-2070 GPU, and it took 3 seconds to load data to/from the GPU and 1 second to solve Poisson learning.
Software Dependencies No The paper mentions 'Python source code' and implies the use of deep learning frameworks (e.g., for autoencoders), but no specific version numbers for Python, libraries, or other software dependencies are provided.
Experiment Setup Yes For MNIST and Fashion MNIST, we used variational autoencoders with 3 fully connected layers of sizes (784,400,20) and (784,400,30), respectively, followed by a symmetrically defined decoder. The autoencoder was trained for 100 epochs on each dataset... In all experiments in this paper, we set µ = 1, Ninner = 40 and Nouter = 20. Steps 9-14 implement the volume constrained projection described in Section 2.4. We set the time step as dτ = 10 and set the clipping values in Step 12 to smin = 0.5 and smax = 2.