Unverified50% confidenceFactExact time
The paper's main result is that for regularly varying innovations (index t ∈ (1,2)), the tail probability of the last iterate is dominated by a single large gradient shock transported through the local Hessian flow.
1
Sources
50%
Confidence
Long-term
Relevance
10/19/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe paper proves that the SGD iterates converge to a stable OU process at the central fluctuation scale and uses this limit to characterize the tail probabilities.66% similarUnverifiedIf a distribution's survival function decays at rate x^{-ν}, it is called a power-law-tailed or heavy-tailed distribution; the smaller ν is, the thicker the tails and the higher the probability of extreme values.63% similarUnverifiedBackpropagation requires caching intermediate activation values from the forward pass to compute gradients, which places a heavy memory burden on very large models.63% similarUnverifiedWhen the factors multiplied by the chain rule are continuously less than 1, gradients decay exponentially (vanishing); when they are continuously greater than 1, gradients grow exponentially (exploding).63% similarUnverifiedStandard Gaussian diffusion models have inherent limitations in capturing heavy-tailed distributions.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1064978API
curl https://kongchang.com/api/v1/knowledge/claims/1064978MCP
get_claim(id=1064978)