The acceptability and adoption of AI are increasing beyond technology companies. Organizations under constant pressure to stay competitive are turning to AI technology to leverage their data resources. AI applications require two essential components: model and data. Historically, most AI solutions have focused on model-centric approaches, relying on vast amounts of data to train models for every scenario. However, businesses are recognizing the limitations of data and are now turning to data-centric AI adoption.

Constraints due to input data volume and quality:

AI adoption is expanding to various industries, including financial services, insurance, manufacturing, logistics and more. Unlike technology companies, most industries lack the luxury of carefully curated data. Therefore, AI systems must address these constraints to be effective. AI solutions must be adaptable to cope with data volume constraints. Consequently, the focus is shifting from ‘Big Data’ to ‘Good Data’ to ensure the effectiveness of AI solutions.

Ability to deal with input data variability:

AI experts concur that data drives success in AI. Training a model for data is often seen merely as a one-time event. Real-world data is frequently noisy and can contain multiple sources of variability such as outliers, irrelevant features, and changing distribution over time. Model-centric AI systems anticipate the data to adhere to certain templates employed during training. Business operations take place in a very complicated environment, with minimal control over the data input. Businesses operate using a network model, interacting with a variety of partners, such as government, regulators and others. The input data cannot be prescribed. It becomes expensive and time-consuming to train the data models with new input data utilising human resources. Any bias that develops during manual labelling frequently causes the AI model to fail. The inconsistent quality of the input data is another issue. The iterative loop of model tuning may cause it to lose priority over time. Inadequate control may result in potential issues with model robustness owing to data drift whenever a critical eye on the data isn’t being deployed.

Client expectations on output data flexibility:

Customers now anticipate predictability and convenience when interacting with brands in today’s digital environment. They also expect these brands to be aware of and responsive to their needs. Businesses need the ability to change their output data in line with clients’ changing needs. Since consumer behaviour is dynamic and ever-changing, the data must also be continually improved to retain model efficacy. A data-centric AI approach facilitates growth when highly predictive models hold the key to a first-rate client experience and adaptability for them.

Data-centric AI approaches can help address inconsistent quality of the input data in several ways:

• Data cleaning and preprocessing: A well designed data-centric AI solution can automate the process of data cleaning and preprocessing to handle missing or inconsistent values and outliers.

• Data quality monitoring: AI models can be trained to identify and flag data points with poor quality, allowing organisations to identify and correct data quality issues in real time. This can be further enhanced with a ‘human in the loop’ approach. 

• Data augmentation: Data-centric AI can automatically generate synthetic data to augment the existing data, making the model more robust for data variability and quality issues.

• Data imputation: AI algorithms can be used to predict missing values in the data, improving the quality and consistency of the input data.

• Data standardisation: AI algorithms can automatically standardise the input data to reduce the impact of scale on the model, making it more robust to variability in the input data.

“Data is food for AI. A data-centric approach allows people in manufacturing, hospitals, farms, to customize the data, making it more feasible for someone without technical training in AI to feed it into an open-source model.” —Andrew Ng, founder and CEO of Landing AI. 

While data is key to the success of an AI solution, organisations may have little control over it as they operate in highly dynamic business environments with multiple stakeholders. In many industries, the conventional approach of training the AI solution with a large volume of training data may not be feasible. Secondly, businesses need to have the ability to deal with input quality issues and be able to improve the quality, considering the limited volume of data. Thirdly, given the complex business environment, business applications need to be adaptable to handle variability of input/output data without having to rely on the retraining of the model. Data-centric AI helps build such adaptive business applications.

This article is written by a member of the AIM Leaders Council. AIM Leaders Council is an invitation-only forum of senior executives in the Data Science and Analytics industry. To check if you are eligible for a membership, please fill out the form here