Image classification tasks—SOTA overview (26th August)

26th August 2024

This is my summary of current landscape in SOTA performances with common image classification tasks.

CIFAR-10 dataset

• size: ?
• leaderboard: https://paperswithcode.com/sota/image-classification-on-cifar-10

Current SOTA trends from paper-with-code:

• 1st Efficient Adaptive Ensamble, 99.612% accuracy
• 2nd and 5th, Vision transformer
• I can see ASAM, an invariant of SAM.
• Is this really updated? SAM 99.7% > Efficient Adaptive Ensamble 99.612%

SAM Paper (to check CIFAR-10 performance back then)

• URL: https://arxiv.org/pdf/2010.01412
• When the paper was published, SAM had sota performances on CIFAR-{10, 100}
• SOTA values are in Table 3
  • Table 1 values are not SOTA. They are to illustrate performance improvements SGD vs SAM.
  • Table 3 has SOTA values, including 99.7% with CIFAR-10 dataset.
• Optimizer was implemented in JAX. Unofficial implementations exist on pytorch and tf.
• Running the sam optimiers reproducible experiment will teach me a lot?
• https://github.com/google-research/sam

CIFAR-100 dataset and sota verification

• https://paperswithcode.com/sota/image-classification-on-cifar-100
• SAM with EffNet-L2 is SOTA in 2020
• From the SAM paper: 100-3.92 = 96.08

ImageNet sota verification

TODO