Limits RegistryPUBLIC RECORD
Limits RegistryKaiming He and colleagues at Microsoft Research introduced deep residual networks ('ResNet') in late 2015, winning the ILSVRC 2015 classification task with a 3.57% top-5 error rate using a 152-layer network — far deeper than prior architectures — made trainable via 'skip connections' that let gradients flow through many layers. It was the first model to beat the human-level ImageNet benchmark Andrej Karpathy had established by hand-labeling images himself in 2014.
The question, scope, and sources behind this Registry record.
Kaiming He and colleagues at Microsoft Research introduced deep residual networks ('ResNet') in late 2015, winning the ILSVRC 2015 classification task with a 3.57% top-5 error rate using a 152-layer network — far deeper than prior architectures — made trainable via 'skip connections' that let gradients flow through many layers. It was the first model to beat the human-level ImageNet benchmark Andrej Karpathy had established by hand-labeling images himself in 2014.
Microsoft Research's ResNet achieved a 3.57% top-5 error rate on the ImageNet classification task in the 2015 ILSVRC, the first model to score below the commonly cited human-level benchmark of approximately 5.1% top-5 error established by a trained human annotator.
Change a parameter to stress-test whether a proposed result is still inside the published specification. This is an audit aid, not a proof checker.
This record has no editable parameters. Read the formal question and assumptions before challenging it.
Current frontiers derived from accepted Claims.
An observed or demonstrated result; no opposing bound is implied.
The frontier is not sacred
Most progress starts with a disagreement that survives contact with evidence. If you can push the known lower bound up or pull the upper bound down, show us the work.
≥ when you have shown that at least this value is achievable.≤ when you have shown that anything above this value is impossible.No vibes. State the value, define the scope, and link the paper, proof, code, or reproduction that lets another person check it. Editors review every challenge before the public record changes.
Challenge this recordAssertions tied to evidence, attribution, and review.
The frontier as it changed over time.
Only accepted Claims matching the current specification contribute to the displayed bounds. Strict inequalities remain open; contradictory Claims require editorial review.
1 accepted Claim, with 1 linked evidence records.
Permanent ID limitsregistry.com/limits/LR-RESNET-HUMAN-LEVEL-IMAGENET
No active verified bounties are linked to this Limit.
View verified bounty tracker ↗No accepted machine-checked reproductions are recorded for this Limit.
Limits Registry. LR-RESNET-HUMAN-LEVEL-IMAGENET. ResNet surpasses estimated human-level performance on ImageNet. 2026.