Big Self-Supervised Models are Strong Semi-Supervised Learners

Authors: Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, Geoffrey E. Hinton

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

Reproducibility Variable Result LLM Response
Research Type Experimental We assess the effectiveness of our method on Image Net ILSVRC-2012 [21] with only 1% and 10% of the labeled images available. Our main findings and contributions can be summarized as follows: ... We combine these findings to achieve a new state-of-the-art in semi-supervised learning on Image Net as summarized in Figure 2.
Researcher Affiliation Industry Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, Geoffrey Hinton Google Research, Brain Team
Pseudocode No The paper describes the Sim CLRv2 framework and its components verbally and through equations but does not include any structured pseudocode or algorithm blocks.
Open Source Code Yes Code and pretrained checkpoints are available at https://github.com/google-research/simclr.
Open Datasets Yes evaluate the proposed method on Image Net ILSVRC-2012 [21].
Dataset Splits Yes While all 1.28 million images are available, only a randomly sub-sampled 1% (12811) or 10% (128116) of images are associated with labels. (...) For fine-tuning... we fine-tune for 60 epochs with 1% of labels, and 30 epochs with 10% of labels, as well as full Image Net labels.
Hardware Specification Yes For pretraining, similar to [1], we train our model on 128 Cloud TPUs, with a batch size of 4096 and global batch normalization [33], for total of 800 epochs.
Software Dependencies No The paper mentions software components like 'LARS optimizer', 'global batch normalization', 'Res Net', and 'Sim CLR' but does not provide specific version numbers for any of them.
Experiment Setup Yes For pretraining... train our model on 128 Cloud TPUs, with a batch size of 4096 and global batch normalization [33], for total of 800 epochs. The learning rate is linearly increased for the first 5% of epochs, reaching maximum of 6.4 (= 0.1 sqrt(Batch Size)), and then decayed with a cosine decay schedule. A weight decay of 1e 4 is used. (...) For fine-tuning... we use a much smaller learning rate, i.e. 0.16 (= 0.005 sqrt(Batch Size)) for standard Res Nets [25], and 0.064 (= 0.002 sqrt(Batch Size)) for larger Res Nets variants (...). A batch size of 1024 is used. Similar to [1], we fine-tune for 60 epochs with 1% of labels, and 30 epochs with 10% of labels, as well as full Image Net labels. For distillation... models are trained for 400 epochs.