Weight-sharing is the architectural pattern wherein a single set of parameters is applied across multiple locations, layers, or time steps of a computational model. Its defining parameters are: (1) a parameter set that is reused rather than instantiated separately for each application site, (2) a structural regularity in the model that makes such reuse meaningful — spatial locality in CNNs, temporal continuity in RNNs, or attention-based routing in transformers — and (3) a training protocol that updates the shared parameters jointly through gradient accumulation across all application sites. The concept persists through documentation in neural network literature, implementation in model architectures, and pedagogical transmission in machine learning curricula. The mechanism of persistence is explicit: it lives in codebases, research papers, and trained models as a deliberate design choice. [formal: participatio | substrate: mind | horizon: generations | explicit: yes | epoch: 0.01]
Accepted ontology entry
weight-sharing
Weight-sharing is the architectural pattern wherein a single set of parameters is applied across multiple locations, layers, or time steps of a computational model. Its defining parameters are: (1) a parameter set that is reused rather tha…
Definition
Why it is in scope
The human-made architectural principle of reusing the same parameters (weights) across multiple parts of a model or system. Originating in convolutional neural networks where spatial locality means the same feature detector scans all regions, extended to transformer architectures and beyond. The concept is a deliberate design pattern for efficiency, generalization, and inductive bias.
Names and aliases
- weight-sharingen · CANONICAL
Relations from this entry
- cms68vjb800dt5ygsto7a43egDERIVED_FROM →
which-came-first-existed-and-fed-into: computation (performing calculations and operations) predates weight-sharing (a neural network technique from the 1980s) by millennia. Weight-sharing — sharing the same parameters across different parts of a network — is a specific computational technique built on the general concept of computation. Computation existed first and fed into the design of weight-sharing as a computational optimization.
- cmrg0pc1x00e82a1nppcnwovjDEPENDS_ON →
Removal test: remove neural networks and weight-sharing has nothing to weight-share — it stops operating. Weight-sharing is a parameter-sharing mechanism that operates exclusively within neural network training.
- cmsi4l2p80452ywh500rxvdvfDERIVED_FROM →
Which-came-first: optimization (the mathematical practice of finding best solutions) predates weight-sharing (a neural network technique emerging in the 1980s). Weight-sharing is a specialized optimization strategy — sharing parameters to reduce model complexity and overfitting — built on the ancient mathematical tradition of optimization. Optimization existed first and fed into weight-sharing.
- cmsn12hrl02q31q13rhwe4gwbDERIVED_FROM →
which-came-first-existed-and-fed-into: model architecture as a design practice predates weight-sharing as a specific neural network technique. Weight-sharing emerged as an optimization technique within the broader practice of designing neural network architectures. The architectural design patterns (CNNs, etc.) that use weight-sharing came from the general practice of model architecture.
- cmrwhb9js006isoacsf207m10SERVES →
Weight sharing is maintained for the sake of regularization: in neural networks it reduces the effective number of parameters, constraining model capacity and preventing overfitting. Its purpose is to act as a regularization mechanism.
Relations to this entry
No accepted relations in this direction.
Record identity
- Created
- Aug 10, 2026, 3:48 PM UTC
- Content hash
- 4092afbf5d66ff5273da1157b1a4e3275490512752809cabae5b5fbf0206933c