On February 20th, NEC announced the industry's first development of two new machine learning operation technologies: one that automatically analyzes and visualizes the cause of degraded prediction accuracy due to AI's inability to adapt to the latest situations during operation based on past data, and another that performs advanced retraining to maintain the accuracy of previously correct data without making it incorrect.
These developments make MLOps (Machine Learning Operations) possible, allowing AI operation while maintaining expected prediction accuracy without the need for data scientists or other experts and without the time and costs. The company plans to utilize the two new technologies in its 'NEC MLOps Service' during fiscal 2023.
AI adoption is progressing as companies accelerate digital transformation (DX), particularly for data analysis. However, changes in social environments during operation can cause a shift in trends from the data learned during development, leading to decreased prediction accuracy. For example, an AI predicting demand in a retail store may face changes when a nearby condominium building is constructed and the number of family‐oriented customers increases.
The concept of MLOps, which integrate development and operation, is gaining attention as a method to prevent these declines in prediction accuracy and continuously maintain and improve the effectiveness of AIs after initial development.
AIs must be monitored to prevent their accuracy from degrading, and if degradation occurs, a cause analysis must be performed before retraining. However, this requires advanced knowledge and significant time and costs.
In response, NEC's newly developed technologies automate general monitoring, a common feature of MLOps tools, and enable advanced retraining that maintains the consistency of correct data without making it incorrect after retraining while also providing cause analysis and visualization.
By utilizing these industry‐first technologies individually or in combination, AI can be operated without advanced knowledge, all while maintaining accuracy and without the time and costs.
The first new technology provides 'automated analysis and visualization of the cause of accuracy degradation during AI operations.' By combining unique indicators such as trend changes in data during learning and operation and the impact on predictions, this technology can automatically analyze the cause of accuracy degradation and present a report with supporting evidence for notable data.
This means that even non‐experts can quickly perform cause analysis for degraded accuracy. The company's demonstration confirmed a 50% reduction in man‐hours needed to analyze the causes of degraded accuracy.
The second new technology is 'advanced retraining, which maintains consistency without making previously correct data incorrect.' Increasing the weight of correct data and retraining enables the accuracy of correct data to be maintained and prevents partial degradation.
This technology is highly effective for retraining imbalanced data that deal with rare events, such as equipment failure prediction and detection of fraudulent financial transactions.
This article has been translated by JST with permission from The Science News Ltd. (https://sci-news.co.jp/). Unauthorized reproduction of the article and photographs is prohibited.