Semi-Supervised Learning via Compact Latent Space Clustering

Authors: Konstantinos Kamnitsas, Daniel Castro, Loic Le Folgoc, Ian Walker, Ryutaro Tanno, Daniel Rueckert, Ben Glocker, Antonio Criminisi, Aditya Nori

ICML 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our approach on three benchmarks and compare to state-of-the art with promising results.
Researcher Affiliation Collaboration 1Microsoft Research Cambridge, United Kingdom 2Imperial College London, United Kingdom 3University College London, United Kingdom.
Pseudocode Yes Algorithm 1 Training for SSL with CCLP
Open Source Code No The paper does not contain any explicit statement about releasing source code or provide a link to a code repository.
Open Datasets Yes We consider three benchmarks widely used in studies on SSL: MNIST, SVHN and CIFAR-10.
Dataset Splits Yes We evaluate on the test-dataset of each benchmark, except for the ablation study where we separated a validation set... For this, we separate a validation set of 10000 images from the training set of each benchmark.
Hardware Specification No The paper mentions 'Tensor Flow GPU implementation' but does not specify any particular GPU model, CPU, or other hardware details used for the experiments.
Software Dependencies No The paper mentions 'Tensor Flow GPU implementation (Abadi et al., 2016)' but does not provide a specific version number for TensorFlow or any other software dependencies.
Experiment Setup Yes In all experiments we used the same meta-parameters for CCLP: In each SGD iteration we sample a batch (XL, y L) DL of size NL = 100... and a batch without labels XU DU of size NU = 100. We use the dot product as similarity metric (Eq. (2)), S =3 maximum steps of the Markov chains (Eq. (9)). LCCLP was weighted equally with the supervised loss, with w=1 throughout training... Exception are the experiments with |DL|=4000 on CIFAR, where lower w=0.1 was used...