Unverified50% confidenceFactExact time
Leaky ReLU outputs a small proportion (usually 1%) of the input value for negative input, thereby avoiding the dead neuron problem.
1
Sources
50%
Confidence
Long-term
Relevance
10/13/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe Φ value of standard feedforward neural networks approaches zero, because information flows only in one direction and lacks recurrent integration.69% similarUnverified精简prompt后往往可以减少30%-50%的输入token而不影响输出质量64% similarUnverified30%–36% of the stable neurons for instruction-tuned generators are concentrated in BERT's layer 12, whereas for base models this proportion is below 14% in all cases.63% similarUnverifiedMany neural networks exhibit an overconfidence calibration bias, where output confidence scores remain close to 1 even when the error rate is high.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1013758API
curl https://kongchang.com/api/v1/knowledge/claims/1013758MCP
get_claim(id=1013758)