Sample- and Parameter-Efficient Auto-Regressive Image Models

https://github.com/elad-amrani/xtra Figure 1. Sample and Parameter Efficiency of XTRA. (Left) XTRA-H/14 (0.6B parameters) outperforms prior state-of-the-art auto-regressive image model (AIM-0.6B [26]) in top-1 average accuracy across 15 diverse image recognition benchmarks, despite being trained on 152× fewer samples. (Right) XTRA-B/16 (85M parameters) outperforms prior auto-regressive image models trained on ImageNet-1k in linear and attentive probing tasks, while using 7-16× fewer parameters.