1. Overview
This document proposes a method for efficiently operating a machine learning model by utilizing modular structure, planned obsolescence, and continuous production-disposal concepts.
2. Modular Model
2.1. Concept
Instead of developing a comprehensive AI model, individual models that perform small-scale judgments are combined and operated.
2.2. Features
•
Dividing into simple judgment units: The AI model does not make comprehensive predictions but performs specific, small judgments (e.g., OK/NO, exceeding/not exceeding the threshold).
•
Combining individual AI models: Various AI models are operated per process and ultimately integrated to perform an overall judgment.
2.3. Expected Effects
•
Reduced data requirements: Since multiple small models are operated instead of one massive model, it is possible to achieve high performance with relatively little data.
•
Lower computational burden: Since it performs simple individual judgments rather than complex predictions, less computation is required.
2.4. Example - Application Method
•
Quality inspection: In product quality inspection, separate models for color classification, strength testing, and defect detection are operated independently, and their results are combined to make a final quality decision.
3. Planned Obsolescence Model
3.1. Concept
A lifespan is set for the model, and normal operation is guaranteed only until that point.
•
Resources (data, computation) are invested only to ensure that the model functions properly within its guaranteed lifespan.
•
Accuracy beyond the model's expiration date is not considered.
3.2. Features
•
Setting an AI model's effective period: A specific lifespan (e.g., 7 days, 14 days, 30 days, until task completion) is set, and once the period ends, the model is discarded.
3.3. Expected Effects
•
Reduced data requirements
•
Lower computational burden
3.4. Example - Application Method
•
Building process models: The state of raw materials is analyzed and trained, and the model is used only until all the raw materials are consumed. When new raw materials arrive, retraining is performed.
4. Continuous Production-Disposal Model
4.1. Concept
The process of discarding models and retraining them with new data is repeated to provide continuous service.
•
Instead of limiting the model's lifespan, new models are continuously trained.
4.2. Features
•
Model disposal and retraining: Instead of maintaining existing models, new models are trained and replaced at each lifecycle.
4.3. Expected Effects
•
Reflecting current data: Since only current data is learned, the model is optimized for the latest conditions.
•
Ensuring continuous service: Even if models have limited lifespans, continuous service can be provided by retraining new ones.
4.4. Example - Application Method
•
Replacing equipment and retraining models: When new equipment is introduced, the data characteristics change significantly, making it necessary to discard the old model and train a new one for optimal performance.
5. Conclusion
The proposed model operation strategy aims to provide an accurate and efficient system by reflecting the current state and environment to generate the most relevant information. The following concepts and approaches are suggested to achieve this goal:
•
Modular Structure
•
Planned Obsolescence
•
Continuous Production-Disposal
