2021/08/30 by Lorenzo Brigato, Björn Barz, Brigato, Lorenzo +5 · 1 citation
Medicine · Computer Science · #COVID-19 diagnosis using AI #Machine Learning and Data Classification #Advanced Neural Network Applications
paper · pdf · doi:10.48550/arxiv.2108.13122
Data-efficient image classification using deep neural networks in settings,\nwhere only small amounts of labeled data are available, has been an active\nresearch area in the recent past. However, an objective comparison between\npublished methods is difficult, since existing works use different datasets for\nevaluation and often compare against untuned baselines with default\nhyper-parameters. We design a benchmark for data-efficient image classification\nconsisting of six diverse datasets spanning various domains (e.g., natural\nimages, medical imagery, satellite data) and data types (RGB, grayscale,\nmultispectral). Using this benchmark, we re-evaluate the standard cross-entropy\nbaseline and eight methods for data-efficient deep learning published between\n2017 and 2021 at renowned venues. For a fair and realistic comparison, we\ncarefully tune the hyper-parameters of all methods on each dataset.\nSurprisingly, we find that tuning learning rate, weight decay, and batch size\non a separate validation split results in a highly competitive baseline, which\noutperforms all but one specialized method and performs competitively to the\nremaining one.\n