Deletion Learning (DL): A General Framework for Enhancing Learning Systems Through Pre-Learning, In-Learning, and Post-Learning Deletion
Abstract
Learning systems are commonly improved by adding data, increasing model capacity, introducing new components, optimizing parameters, or strengthening existing mechanisms. However, these approaches do not exhaust the possible directions for improvement. The removal of selected objects may also improve a learning system by reducing redundancy, eliminating interference, simplifying computation, changing learning dynamics, or removing components that contribute insufficiently to a specified objective. This paper proposes Deletion Learning (DL), a general conceptual framework that treats deletion as an explicit design and optimization operation across the lifecycle of learning systems. Unlike approaches restricted to the removal of training data, learned information, or neural network parameters, DL encompasses a broad range of potentially removable objects, including data, information, parameters, structures, connections, representations, rules, objectives, behavioral mechanisms, computational procedures, memory entries, and entire subsystems. The framework distinguishes three principal stages: pre-learning deletion, in-learning deletion, and post-learning deletion. It further organizes deletion according to target category, operational mechanism, decision policy, reversibility, and intended effect. A formal model is introduced to characterize deletion operations, estimate their marginal contribution to system objectives, represent interactions among multiple deletions, and incorporate capability-preservation constraints. The framework also proposes a systematic research methodology for evaluating deletion across different learning paradigms, including supervised learning, self-supervised learning, reinforcement learning, adaptive computation, and multi-component intelligent systems. The central proposition is that appropriately selected deletion operations can improve specified properties of a learning system when the benefits of removal exceed its costs and essential functionality is preserved. DL does not assume that deletion is universally beneficial or that all candidate objects can be independently removed. Instead, it provides a general framework for investigating what should be deleted, when deletion should occur, how removal should be implemented, and how its consequences should be evaluated. This framework establishes deletion as a potentially important dimension of learning-system design and optimization.
Related
- Deletion Learning for Multi-Agent Systems (DL-MAS): A General Framework for Enhancing Collective Intelligence Through Pre-Training, In-Training, and Post-Training Deletion
- Enhancing Large Language Models Through Pre-Training, In-Training, and Post-Training Deletion
- A Conditional, Boundary-Aware, and Auditable Deletion-Effect Weighting Framework for MCDM with Positive Decision Matrices and a Fixed Set of Alternatives
- Deletion-Driven 3D Generation: A Progressive Removal Framework for Shape, Topology, and Material-Aware Three-Dimensional Synthesis
- HGPRGDel: GPR-Aware Learning and Ranking for Growth-Coupled Gene Deletion Design
BibTeX
@misc{zhang2026deletion,
title = {{Deletion Learning (DL): A General Framework for Enhancing Learning Systems Through Pre-Learning, In-Learning, and Post-Learning Deletion}},
author = {Jincheng Zhang},
howpublished = {Zenodo},
year = {2026},
month = oct,
doi = {10.5281/zenodo.23464348},
url = {https://zenodo.org/records/23464348},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Deletion Learning (DL): A General Framework for Enhancing Learning Systems Through Pre-Learning, In-Learning, and Post-Learning Deletion
AU - Jincheng Zhang
PY - 2026
DA - 2026/10/19
PB - Zenodo
DO - 10.5281/zenodo.23464348
UR - https://zenodo.org/records/23464348
AB - Learning systems are commonly improved by adding data, increasing model capacity, introducing new components, optimizing parameters, or strengthening existing mechanisms. However, these approaches do not exhaust the possible directions for improvement. The removal of selected objects may also improve a learning system by reducing redundancy, eliminating interference, simplifying computation, changing learning dynamics, or removing components that contribute insufficiently to a specified objective. This paper proposes Deletion Learning (DL), a general conceptual framework that treats deletion as an explicit design and optimization operation across the lifecycle of learning systems. Unlike approaches restricted to the removal of training data, learned information, or neural network parameters, DL encompasses a broad range of potentially removable objects, including data, information, parameters, structures, connections, representations, rules, objectives, behavioral mechanisms, computational procedures, memory entries, and entire subsystems. The framework distinguishes three principal stages: pre-learning deletion, in-learning deletion, and post-learning deletion. It further organizes deletion according to target category, operational mechanism, decision policy, reversibility, and intended effect. A formal model is introduced to characterize deletion operations, estimate their marginal contribution to system objectives, represent interactions among multiple deletions, and incorporate capability-preservation constraints. The framework also proposes a systematic research methodology for evaluating deletion across different learning paradigms, including supervised learning, self-supervised learning, reinforcement learning, adaptive computation, and multi-component intelligent systems. The central proposition is that appropriately selected deletion operations can improve specified properties of a learning system when the benefits of removal exceed its costs and essential functionality is preserved. DL does not assume that deletion is universally beneficial or that all candidate objects can be independently removed. Instead, it provides a general framework for investigating what should be deleted, when deletion should occur, how removal should be implemented, and how its consequences should be evaluated. This framework establishes deletion as a potentially important dimension of learning-system design and optimization.
ER -