Verified70% confidenceSolutionExact time
Numerical gradient checking approximates gradients using the central difference formula (f(θ+ε) - f(θ-ε)) / (2ε) and compares them with analytical gradients; a relative error below 1e-5 is considered correct, while an error above 1e-3 almost certainly indicates a bug.
4
Sources
70%
Confidence
Long-term
Relevance
8/7/2026
First Seen
Sources
picodl：用纯NumPy从零手写深度学习库的技术解析
Related Claims
UnverifiedNumerical gradient checking approximates the derivative via finite differences: for each parameter θ, compute [L(θ+ε) - L(θ-ε)] / (2ε), where ε is a very small number (such as 1e-4).77% similarUnverifiedFor constant mappings and a constant-mapping prior, the exponential base of the identification regret is exactly β_★ (about 3.583), rather than the 4 suggested by the moment range; for monotone mappings, this base lies in the interval [2, β_★].67% similarVerifiedFP8 has two variants: E4M3 (4-bit exponent + 3-bit mantissa), which has higher precision and is suited for forward inference, and E5M2 (5-bit exponent + 2-bit mantissa), which has a larger dynamic range and is suited for gradient computation.66% similarUnverifiedThe GK summary (Greenwald-Khanna) uses counters with error bounds to provide deterministic approximate guarantees for quantiles.65% similarUnverifiedSingle-turn tests use strict exact-parameter scoring weighted by severity (L4=8 points decreasing to L0=0 points).65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/705532API
curl https://kongchang.com/api/v1/knowledge/claims/705532MCP
get_claim(id=705532)