Successful deployment of AI practice model
The deployment of artificial intelligence is inseparable from the injection of large-scale high-quality training data. With the development of artificial intelligence, the management system of big data is becoming more and more perfect. Data collection and governance are two complicated factors when enterprises make AI deployment strategies. There are several best practices that can serve as models for building and deploying effective AI solutions. Establishing a long-term and comprehensive AI governance framework (especially around data governance) and scalable data processes will also be a necessary process for enterprises to build AI solutions. This article will break down the main considerations of AI governance and guide the creation and maintenance of training data processes step by step .
Defining AI Governance
AI governance is the framework for overseeing the use and implementation of AI in an enterprise. Different companies are influenced by their industry, internal regulations, regulatory requirements and local laws on how to define this framework. In any case, there is no one-size-fits-all approach; each business should choose the option that best suits its needs. Generally speaking, three key areas of AI governance are usually included in the framework:
performance
How to measure the performance of the model is an important factor in the development process. The development team should develop a set of metrics to track from initial model build to post-deployment to ensure the model performs (continues to execute) as expected. The above indicators need to include several key factors: Accuracy For accuracy, on the one hand, the precision and recall of the model need to be considered. Does the prediction meet the desired confidence threshold? If the answer is no, iterations are required. On the other hand, also consider whether the model has the context to accurately predict the context needed. This is where the data will tell the answer, but make sure it includes all use cases and known edge cases. Bias/Fairness incorporates metrics that measure bias in model performance. Third-party tools are currently available to help track this metric. Bias can come from sampling (i.e. how, from where and by whom the data is collected) or from data annotators. For example, top facial recognition software has shown that darker-skinned people have higher error rates than lighter-skinned people. For example, black women had a misidentification rate of more than 25 percent, compared with just 1 percent for white men. The problem is the data being collected (underrepresented by people of color) and the people annotating the data (mostly white), the lack of diversity leads to less than ideal solutions. Implementing best practices in AI data acquisition and governance frameworks can reduce bias in AI .
transparency
According to relevant laws and regulations, companies are usually required to demonstrate how AI models make decisions. The General Data Protection Regulation (GDPR) is an example of the EU's right to transparency for consumers. Even without regulation, the interpretability of AI models remains critical to end users and reproducibility. As you build your model, fully document how it works. A governance framework can address documentation practices and a commitment to transparency.
ethical standards
Ethical standards are a third area that is common in AI governance frameworks. Ethical standards play a role throughout AI implementation, first to ensure that the intent of the solution is ethical, and finally to ensure that the model continues to function as intended. In this section, define what responsible AI looks like from pilot to production, and what processes will be used to ensure that needs are met. Companies embarking on the data journey and leveraging the business value of data across the information supply chain will need to navigate the challenges of self-data capture services analysis. And the criticality of metadata management and data catalogs cannot be diminished. We will introduce the artificial intelligence data management requirements in detail in the next article , click to read.