The Spark That Drives Machine Learning to Shine

Technology | Jon Hanzelka |
14 September 2023

The power of machine learning (ML) is in the details. When combined with data, ML is a convenient, context-fuelled resource that can push human efficiency and productivity to new heights.

The fuel, in this case, is synthetic data, which we define as, “artificially generated machine learning training data that mimics the characteristics of real-world phenomena.” The positives of leveraging this information are not lost on decision-makers. According to Gartner, synthetic data is expected to completely overshadow real data in artificial intelligence (AI) models by 2030, with some believing “you won’t be able to build high-quality, high-value AI models without it.” However, with empirical insights so critical to the success of ML, a lack of a coherent data strategy could stop that momentum in its tracks.

Finding the footing to overcome those challenges can be key to organisations unlocking the true value of this innovative and impactful solution. Could synthetic data be the missing piece to pushing these ML integrations over the top?

How machines learn

As an analogy to understand how ML works, consider its closest counterpart: the human brain. Humans learn and retain information through repeated experiences and feedback that refines our knowledge of the world around us.

ML operates in a comparable manner through a process known as supervised learning. In this instance, the human brain is replaced by a neural network. By providing the neural network with a curated input of imagery and corresponding annotations that describe the images, the network learns to recognize patterns that can then be applied to new, similar data.

In our rapidly digitizing world, computers are becoming more than passive presenters of images or video. Through computer vision, they have become insightful interpreters, providing an understanding of the visual world around us, deriving value from aspects like:

Classification: The network assesses the contents of the image and categorizes it into predefined labels, or categories.

Detection: The network learns to identify the general region or bounding box for where these objects exist in the image.

Segmentation: The network defines a polygon encompassing an object’s shape or boundary with more detail and precision than a bounding box. The tighter fit around the object keeps less relevant background information out of the annotation, leading to more detail for complex shapes.

Understanding: The network can identify specific intrinsic properties of an object, such as the presence of a defect.

Targeting: The network can comprehend an object’s position within 3D space.

Tracking: The network can interpret an object’s motion properties.

Leveraging these capabilities to solve complex problems in real-world applications is often gated by having access to the right data. Inaccuracies inherent in manually labelled data can lead to cost overages; this illustrates the importance of every industry having ground truth data that is accurate and precise, especially in areas like healthcare. Misinformed data can yield unreliable predictions that make the adoption of machine learning solutions in a production environment unrealistic.

Synthetic data can account for these inaccuracies and help to address many of the most challenging aspects of collecting data for training ML applications.

How synthetic data troubleshoots machine learning challenges

First, it is important to properly contextualize the term synthetic data because it can often get lumped in as a generative AI offshoot. With synthetic data, real-world data is augmented or replaced with imagery generated in 3D applications using techniques common to the visual effects and gaming industries. Generative approaches and models can also be factored into the process and leveraged to complement workflows within solutions, when applicable.

The composition of these photo-realistic synthetic images is informed by real-world variables and the requirements of the machine learning application. These parameters capture the diversity that is encountered in the real world in an unbiased manner which can be difficult to achieve when only using real data. Having this granular control over the data that is fed into the machine learning process can address many commonly encountered issues, such as:

Rare data: Consider safety and inspection, an industry heavily reliant on data that may not exist in the quantities needed to train an ML system. Synthetically generated imagery of these rare cases can be produced to fill in these gaps in real data. This enables ML vision systems to be trained in ways that would not otherwise be possible.

Privacy protection: ML’s reliance on data can endanger personal privacy. For some industries, obtaining the data needed to build models can be a long and complicated process because it first needs to go through cleanup or anonymization, as it might contain PII (Personally Identifiable Information), medical history or other sensitive information. Synthetics allows the creation of imagery or other types of data that have no association with real individuals.

Data precision: When dealing with millions of data points, human error can rear its head and lead to a mislabelled sample or errant figure. With synthetic data, pixel-perfect annotations are created so ML can remain efficient and unlock a host of capabilities.

Previsualization: Often in a hardware production cycle, physical devices may not exist when the data is needed. Synthetic data can be an early means of producing accurate representations of the output of future devices. This enables the unblocking of algorithmic development and can even serve to better inform the hardware prototyping phase.

Reproducibility: Synthetic data can be replicated on demand. This means that as project specifications evolve, training data can adapt to the latest requirements. This reproducibility also enables targeted parameter refinement to produce the most effective training set for a scenario.

Generalization: This concept hinges on the idea that creating a robust, flexible model means training it on an unbiased collection of data that represents the diversity that may be encountered. For example, if data is only collected from a single geographic region, contextual insights could be too biased to provide a truly global representation. With synthetic data, explicit control over distribution parameters allows for the negation of unwanted bias.

Putting ML efficiency in the hands of a trusted partner

Investing in reliable machine learning can represent a significant step forward toward a more intuitive and proficient method of production. Every industry seeks ways to increase expediency and cost-effectiveness without sacrificing quality, and ML can be a data-driven means to that end.

Want to take a deeper look into our process for creating synthetic data? Click here to watch a video of Jacob Berrier and I presenting “Beyond Visible Light: Generating Synthetic Data in Unique Spectrums” at SIGGRAPH 2023.

And if you are ready to turn data simulations into details that power real-world solutions, contact us here.

Jon Hanzelka

TECHNICAL ART DIRECTOR

Jon Hanzelka is a Technical Art Director at Endava where he oversees the synthetic pipeline development and client engagement that enables machine learning solutions across a broad range of industry verticals. He has worked in tech and entertainment for over 18 years, contributing to films such as, "The Avengers," "Inception" and "The Dark Knight Rises" and helping ship products like Microsoft HoloLens v1 and 2. For the past 8 years, he has been focused on the field of synthetic data and how it can be used to solve many of the industry’s most challenging machine-learning problems.