Iterative Teaching by Label Synthesis

Authors: Weiyang Liu, Zhen Liu, Hanchen Wang, Liam Paull, Bernhard Schölkopf, Adrian Weller

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, we empirically demonstrate the value of our framework. and This section comprehensively evaluates LAST in the omniscient teaching scenario. Experimental details and more results (including BLAST) are given in Appendix G and Appendix F, respectively.
Researcher Affiliation Academia 1University of Cambridge 2MPI for Intelligent Systems, Tübingen 3Mila, Université de Montréal 4CIFAR AI Chair 5The Alan Turing Institute
Pseudocode Yes Algorithm 1 Omniscient Greedy LAST Initialize t=1, w0, ϵ and T; while wt w 2 ϵ or t T do Randomly select a sample xt from the pool; Solve Eq. (1) to synthesize the label yt: Use the synthesized label yt for the update: wt = wt−1 − ηt ℓ(xt, yt|wt−1) Set t ← t + 1; end
Open Source Code No The paper does not contain any explicit statement about releasing source code or a link to a code repository.
Open Datasets Yes For real image data, we use 3/5 digits in MNIST.
Dataset Splits Yes For synthetic regression data, we generate 1000 data points (500 for training and 500 for testing)... For linear logistic regression, we generate 500 two-dimensional data points (300 for training and 200 for testing). and For the binary classification on MNIST (digit 3 and 5), we randomly select 500 samples (250 for digit 3 and 250 for digit 5) from the original MNIST training set as our training data, and 200 samples (100 for digit 3 and 100 for digit 5) from the original MNIST test set as our test data.
Hardware Specification No The paper does not specify any particular hardware (e.g., GPU/CPU models, memory, cloud instances) used for running the experiments.
Software Dependencies No The paper mentions 'All algorithms are implemented with Python and PyTorch' but does not specify version numbers for these software components or any other libraries.
Experiment Setup Yes The learning rate is selected from {0.001, 0.005, 0.01, 0.05, 0.1, 0.5} via grid search for all algorithms and the one that gives the best performance in terms of convergence is selected. The batch size is 1 for all algorithms. For MLP learners, we use a 2-layer MLP with 100 hidden neurons. The weights are initialized by Kaiming Initialization.