Code to reproduce results in "How Well Do Unsupervised Learning Algorithms Model Human Real-time and Life-long Learning?" (to appear in NeurIPS 2022 Dataset and Benchmark Track).
In this work, we build two benchmarks to compare the real-time and life-long visual learning dynamics between unsupervised algorithms and humans. The two folders are correspondingly for each benchmark.
Check the README at the life_long
folder. This benchmark needs the SAYCam dataset for training and the ImageNet dataset for testing to fully reproduce the results.
Per-checkpoint performance of all algorithms under all learning conditions can be found at link. The algorithms and training conditions are in the same order as Fig. S1. The algorithm names should be self-explanatory.
Check the README at the real_time_related
folder. This benchmark provides the pretrained models (see the README at real_time_related/real_time_test/pytorch_scripts
) but still needs the ImageNet and VGGFace2 datasets (see instructions in the same README about how to download this dataset).