Unverified50% confidenceFactExact time
FP16 has only 5 exponent bits, and the smallest representable normal number is about 6×10^-5, so input gradients in backward propagation easily underflow to zero, causing shallower layers to learn nothing at all.
1
Sources
50%
Confidence
Long-term
Relevance
10/20/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedSigmoid函数的导数最大值仅为0.25，经过数十层连乘后梯度会指数级衰减67% similarUnverifiedUnder the zero-shot setting on the TurnBench development set, FD-VAD achieves a maximum EOT recall of 0.853 under the constraint of false positive rate FP≤0.10.65% similarUnverifiedIn the setting closest to reality (where unstated constraints must be actively discovered), the best-performing model configuration achieves an average FDE-Gym score of only 0.56.63% similarUnverified4-bit quantization compared to FP16 typically results in a 0.5–2% drop in benchmark accuracy.63% similarUnverifiedQ8_0 quantization compresses model weights from 32-bit floating-point numbers to 8-bit integers, with precision loss typically within 1% compared to FP32.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1074635API
curl https://kongchang.com/api/v1/knowledge/claims/1074635MCP
get_claim(id=1074635)