Unverified50% confidenceConceptExact time
Camerata proposed the concept of "recursive generalization": neurons branch and grow along related directions during training, and each branch has a provenance rather than being forcibly stuffed in as superposition.
1
Sources
50%
Confidence
Long-term
Relevance
10/17/2026
First Seen
Sources
Neural Networks Were Always Clean? Goodfire Researcher Overturns Superposition Theory
bilibili迦太基盐商10/16/2026
Related Entities
Related Claims
UnverifiedNeural networks exhibit permutation symmetry—reordering the neurons in a given layer leaves the network's functionality completely unchanged, as long as the weights of adjacent layers are adjusted accordingly.67% similarUnverified赫布定律指出一起激活的神经元连接在一起，长期不被激活的突触连接会在突触修剪过程中被清除63% similarPartially VerifiedThe self-attention mechanism of the Transformer resolves the long-range dependency problem, while its parallelized architecture resolves the training efficiency problem.63% similarUnverifiedSAC's entropy regularization mechanism helps agents in high-dimensional continuous action spaces maintain exploration during training and avoid prematurely converging to rigid, monotonous action patterns.63% similarUnverifiedThe weakening of pattern separation ability during aging tilts the balance toward pattern completion, which is the neurocomputational basis for memory confusion.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1052471API
curl https://kongchang.com/api/v1/knowledge/claims/1052471MCP
get_claim(id=1052471)