Unverified50% confidenceFactExact time
In Google Brain's 2020 ViT paper, it was noted that ViT trained from scratch on ImageNet-1K (1.28 million images) underperforms CNNs of equivalent scale, but begins to surpass CNNs when pre-training data is scaled to ImageNet-21K (14 million images) or JFT-300M (300 million images)
1
Sources
50%
Confidence
Long-term
Relevance
8/14/2026
First Seen
Sources
小数据集训练DETR：无人机枯树检测实战解析
Related Claims
Unverified残差连接支撑了152层超深网络的训练，并在2015年ImageNet竞赛中以低于4%的错误率首次超越人类水平（约5%）68% similarUnverifiedA 2024 study indicates that even when AI-generated content accounts for only 10–20% of training data, measurable quality degradation can occur after multiple generations of training.65% similarUnverifiedGoogle Brain团队于2022年系统记录了涌现能力现象，发现某些能力在模型达到特定规模前几乎为零，之后突然跃升63% similarUnverifiedA single AI inference request consumes little electricity, but because requests number in the billions, the cumulative energy consumption may exceed that of training itself.62% similarUnverifiedAnthropic accuses Alibaba of using 29 million interactions with Claude to distill training data for training the Qwen series of models.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/745157API
curl https://kongchang.com/api/v1/knowledge/claims/745157MCP
get_claim(id=745157)