Weighted Distillation with Unlabeled Examples

Authors: Fotis Iliopoulos, Vasilis Kontonis, Cenk Baykal, Gaurav Menghani, Khoa Trinh, Erik Vee

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

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate significant improvements on popular academic datasets and we accompany our results with a theoretical analysis which rigorously justifies the performance of our method in certain settings.
Researcher Affiliation Industry Fotis Iliopoulos Google Research fotisi@google.com Vasilis Kontonis Google Research kontonis@google.com Cenk Baykal Google Research baykalc@google.com Gaurav Menghani Google Research gmenghani@google.com Khoa Trinh Google Research khoatrinh@google.com Erik Vee Google Research erikvee@google.com
Pseudocode Yes We describe our method below and more formally in Algorithm 1 in Appendix A. The pseudocode for our method can be found in Algorithm 2 in Appendix A.
Open Source Code No The paper states 'Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [No]' in Section 6, point 3a.
Open Datasets Yes SVHN [34] is an image classification dataset... CIFAR-10 and CIFAR-100 [24] are image classification datasets... Celeb A [15] is a large-scale face attributes dataset... Image Net [41] is a large-scale image classification dataset...
Dataset Splits Yes Our experiments are of the following form. The academic dataset we use each time is first split into two parts A and B. Part A, which is typically smaller, is used as the labeled dataset Sℓwhere the teacher model is trained on (recall the setting we described in Section 2.3). Part B is randomly split again into two parts which represent the unlabeled dataset Su and validation dataset Sv, respectively. ... In each experiment we use the first N {7500, 10000, 12500, 15000, 17500, 20000} examples as the labeled dataset Sℓ, and then the rest 73257 N images are randomly split to a labeled validation dataset Sv of size 2000, and an unlabeled dataset Su of size 71257 N. ... We use a validation set of 2000 examples.
Hardware Specification Yes We implemented all algorithms in Python making use of the Tensor Flow deep learning library [1]. We use 64 Cloud TPU v4s each with two cores.
Software Dependencies No The paper mentions 'Tensor Flow deep learning library [1]' but does not specify a version number for TensorFlow or any other software dependencies.
Experiment Setup Yes We always choose the temperature in the softmax of the models to be 1 for simplicity and consistency, and our metric for confidence is always the margin-score. Implementation details for our experiments and additional results can be found in Appendices B and C. ... In each experiment we use the first N {7500, 10000, 12500, 15000, 17500, 20000} examples as the labeled dataset Sℓ, and then the rest 73257 N images are randomly split to a labeled validation dataset Sv of size 2000, and an unlabeled dataset Su of size 71257 N.