Unverified50% confidenceFactExact time
Global norm clipping must be executed only after all gradients are computed, because the global norm needs to be aggregated first, which inherently conflicts with the "release immediately after computation" of optimizer-in-backward.
1
Sources
50%
Confidence
Long-term
Relevance
11/3/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedGradient clipping works by proportionally scaling down the entire gradient vector when its norm exceeds a preset threshold, so that its norm equals exactly that threshold.79% similarUnverifiedThe method treats pruning as a combinatorial optimization problem: selecting a subset from all partial solutions to maximize the probability that the subset, after continued generation, contains at least one correct solution.77% similarUnverifiedGradient accumulation is crucial in automatic differentiation: when a variable influences the loss function through multiple computational paths, the gradients from each path must be summed72% similarUnverifiedThe reduced ancestor set compresses the graph traversal scale for a single gradient computation to polylogarithmic scale by truncating and retaining only a finite-depth ancestor set, with the bias introduced typically being controllable in practice.70% similarUnverifiedOccupancy-measure-based algorithms rewrite the policy optimization problem as convex optimization over the occupancy measure polytope, thereby globally controlling error accumulation across all state-action pairs, which is the fundamental reason their regret bound has one fewer H factor.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1130236API
curl https://kongchang.com/api/v1/knowledge/claims/1130236MCP
get_claim(id=1130236)