1. Insights
  2. AI Data
  3. Article
  • Share on X
  • Share on Facebook
  • Share via email

The contemporary trifecta of data-centric AI

Posted July 5, 2022
Robotic hand interacting with an interface meant to represent artificial intelligence

In the last few decades, we’ve witnessed the widespread emergence of artificial intelligence (AI) in several industries. AI’s tremendous success can be attributed mainly to machine learning (ML) innovation fueled by the model-centric AI approach. The primary focus of this approach is to continuously improve and iterate the code to improve model performance, whereas the training data is a fixed component throughout the innovation process. However, the focus is now shifting to data-centric AI, where the emphasis is on the data fed into ML models.

What is data-centric AI?

Mature deep learning algorithms and advanced neural network architectures have paved the way for the paradigm shift to data-centric AI. Here, the code is the fixed component, and the data fed into the code is re-engineered to help models produce better outputs. The data is typically enhanced via improving labeling consistencies, removing noisy data and conducting in-depth error analysis, along with a number of other data augmentation techniques compatible with the problem at hand.

The revised approach has proven successful in practicality. The 2021 research paper titled, A Data-Centric Approach to Design and Analysis of a Surface-Inspection System Based on Deep Learning in the Plastic Injection Molding Industry, found that the method of fine-tuning the data, instead of tweaking the code, is more effective in improving model performance. Using different data pre-processing techniques to overcome malfunctions and maintaining data consistency resulted in improved inspection accuracy.

The indispensable role of data in the AI lifecycle

With the advent of data-centric AI, functions such as data sourcing, engineering, annotation and validation ‒ previously considered pre-processing activities ‒ are now critical to ensure model enhancements.

Diagram depicting the indispensable role of data in the AI lifecycle

For example, a smartphone manufacturer building a computer vision model that detects screen damage could experiment with two AI approaches. With a model-centric approach, the emphasis is to improve the baseline performance of the damage detection model by enhancing the code. This often results in minimal performance improvement in each iteration. However, in a data-centric approach, the focus shifts to collecting diverse, high-quality data and reducing labeling inconsistencies via precise labeling parameters, improved human-in-the-loop consensus, detailed error analysis and other data re-engineering methods.

In principle, the latter approach is believed to have a significant impact on the model’s outcomes due to the improved data diversity, label consistency, accuracy and dataset volume. Models tend to perform better when they are ingested with data iteratively and have a continuous supply of high-quality data. More intricate AI systems, such as self-driving cars, search engines and recommender systems, are data-hungry and require highly enriched datasets to produce better machine learning outcomes. But how does one navigate the challenges of adopting the data-centric AI approach?

Data challenges impeding the development of data-centric AI

The significant bottleneck of AI is no longer ML models but the data required. A lack of high-quality data can significantly derail initiatives and slow down AI progress. Collecting, cleaning, labeling and aggregating data for training, testing and validating models require cumbersome human effort. As a result, these data processing activities can also be expensive and time-consuming, and can be an immense challenge for ML teams. Industry expertise is also essential for ML teams to outline detailed labeling parameters, configurable tools to ensure labeling consistency and reliable metrics to measure and evaluate label accuracies. Moreover, training, qualifying and managing several annotators for a project can quickly become a complicated task.

Other challenges faced by ML teams include:

  • Slower implementations due to increased reliance on manual labeling.
  • Overcoming AI bias due to skewed data or inaccurate or biased labeling.
  • Managing data project configurations and automating data workflows.
  • Managing workforce training, evaluation, task assignments, etc.
  • Building custom tool features for varied edge cases across different industries.
  • Defining quality metrics and benchmarks and designing efficient evaluation processes.
  • Creating thorough labeling guidelines to maintain labeling consistency across all annotators.
  • Maintaining data privacy and security to comply with data governance regulations.

The essential guide to AI training data

Discover best practices for the sourcing, labeling and analyzing of training data from TELUS International, a leading provider of AI data solutions.

Download the guide

The contemporary trifecta of data-centric AI: Platform, professionals and processes

The right combination of platform, professionals and processes can help close the AI data gap. ML engineers and AI innovators can maximize the benefits of high-quality data and improve model outputs by combining the power of an advanced labeling platform with highly-trained experts and a systematic data labeling process.

A sophisticated AI training platform

A platform with data labeling automation capabilities helps ML teams achieve improved productivity rates, reduce human effort and gain higher cost savings when re-engineering data to suit their models. Sophisticated quality control tools can accurately measure and evaluate label quality and help ML teams adopt an analytics-driven approach to fine-tune training datasets. Built-in data security and privacy controls in an AI training platform are prerequisites to ensure compliance with all data governance regulations. Moreover, automated workflows and data management systems support a continuous data supply chain to train, test and validate ML models. For example, Ground Truth Studios, TELUS International’s AI Training Platform, offers robust configurability and automation that simplifies the implementation of a data-centric AI approach.

A diverse and well-trained professional ecosystem

A diverse and well-trained team of experts injects the necessary human judgment in the data-centric AI approach while combating potential AI biases. A diverse crowd of annotators helps to overcome the over/under representation of certain identities within ML datasets, along with inconsistencies and label discrepancies. At TELUS International, we combine human intelligence with technology to help ML teams build successful AI. Teams can leverage our platform, Ground Truth Studios, which also serves as a central hub for managing, supporting, training and engaging the AI Community dedicated to AI projects.

Carefully-curated processes designed for success

Along with supercharging the human-in-the-loop process with a powerful platform, end-to-end project management ensures that data flows seamlessly from one touchpoint to the next. Streamlining data processing activities include enforcing data security and handling protocols, creating detailed annotator training instructions and evaluation standards for various projects, assigning data-labeling tasks to a diverse workforce, re-engineering datasets to accommodate edge cases or complex scenarios and much more. A well-functioning data-delivery process is key to building high-quality datasets for successful AI systems.

Innovators are constantly pushing the boundaries of AI applications and the data-centric approach signifies a promising future advancement. Leverage TELUS International’s intelligent mix of platform, professionals and processes to supercharge your AI data pipelines. Learn how we can help you today.

Check out our solutions

Enrich your data with our range of human-annotation services at scale.

Learn more