Interoperability and the Future of Machine Learning

Author

Robert Koch

I write about AI, SEO, Tech, and Innovation. Led by curiosity, I stay ahead of AI advancements. I aim for clarity and understand the necessity of change, taking guidance from Shaw: 'Progress is impossible without change,' and living by Welch's words: 'Change before you have to'.

Artificial Intelligence and its subset Machine Learning are at the heart of innovation for digitally transformed businesses. However, ML, in particular, needs to be highly interoperable for smart technologies to be truly disruptive and innovative at scale.

If there was no interoperability, AI development would be limited and only accessible to big tech. This is because only tech giants have access to the necessary resources and, more importantly, the most data that makes continuous and meaningful learning possible.

Table of Contents

• What is Data Interoperability?

• Why is Data Interoperability Important for ML Development?

• Why is Interoperability Critical to Healthcare?

• Conclusion

What is Data Interoperability?

In software development, the term “interoperability” describes the ability of two different systems to communicate with each other seamlessly. As a key characteristic of future ML development, interoperability will be vital to application development for industries like banking, finance, healthcare, and much more.

The idea here is to make disparate data work together and deliver enhanced user experiences. After all, we all have hundreds of thousands of data points related to everything we do in our daily lives. So, why not leverage it for the maximum benefit of the user/customer?

More About Data Interoperability?

When tech giants tried to control this space with lock-in models, it had a negative impact on AI and ML development. Smaller companies couldn’t compete, and developers were also tied to services and providers like Amazon’s AWS. Whenever this happens, there’s a real risk of missing out on potentially highly robust AI architecture developed by much smaller companies.

For example, Google’s TensorFlow is one of the most popular AI frameworks because of its high computational power. However, because it lacks several pre-trained AI models, it’s not the best option when it comes to innovation and accelerating time to market.

Similarly, Amazon’s AWS boasts both comprehensive data analysis tools and a high level of security but falls short when it comes to flexibility with specific ML algorithms. In this scenario, it’ll be much easier for development teams to use the best frameworks and several related features if they are all interoperable and flexible. If not, they would have to keep switching cloud services providers.

Some of the most innovative AI developments emerged out of smaller startups (like Nauto), so it doesn’t make sense to continue with lock-in models that make it challenging to switch AI frameworks or use them across different AI architectures.

Tip:

There is no reason to commit to one of these lock-in models. Create your own comprehensive and representative ML Training Datasets with our help.

Datasets for Machine Learning & AI

Why is Data Interoperability Important for ML Development?

Data interoperability is now a critical component of AI and ML development. This is because it helps level the field and allows smaller providers to access data resources once only available to tech giants.

Interoperability between and across different platforms also helps intelligent algorithms formulate a shared understanding and helps make other AI and ML models across vendors more flexible. This approach helps accelerate innovation and helps advance the AI field.

The good news is that tech giants like Facebook and Microsoft launched the Open Neural Network Exchange (ONNX) approximately five years ago to enable the seamless transfer between AI frameworks and ML models.

This event was essentially groundbreaking as developers no longer had to commit to a specific AI framework. They could simply start their research and quickly move between different tools and combinations. For example, they could mix and match technologies like Cognitive Toolkit (Microsoft) and Caffe2, Pytorch (Facebook).

Although Amazon wasn’t part of this initiative, they did pilot a voice interoperability project with the likes of Facebook, Garmin, and Xiaomi. In this case, the company was able to ensure compatibility between systems to allow voice services to work across platforms seamlessly.

Why is Interoperability Critical to Healthcare?

For the healthcare sector, in particular, interoperable ML tools have a real potential to enhance and, in some cases, save lives. In a way, it’s like an expert doctor working together with a souped-up robot doctor to provide patients with the best care possible.

Interoperability will potentially optimize and accelerate the diagnosis of critical illnesses; it could also help healthcare services providers improve care, lower costs, deliver enhanced patient experiences, and reduce the risk of medical error (to name a few).

For example, medical procedures like Computerized Tomography (CT) scans generate a massive amount of data about a single patient. However, this information is different from what doctors actually enter into their proprietary databases during a routine check-up. By quickly and automatically integrating these two types of data for analysis, doctors will be able to diagnose potentially critical illnesses rapidly and accurately.

Constraints of Interoperability in the Current Healthcare Sector

• Things didn’t exactly go to plan as healthcare services providers are under constant pressure to accelerate their infrastructure updates driven by recent legal and regulatory changes. For example, healthcare providers must ensure the security and privacy of the sensitive data stored in their databases (and that in itself is no easy feat!).

• It’s also increasingly challenging to make it work on a global scale because of a lack of communications standards (across different patient records systems), the absence of patient identification across various health data exchanges, poor data sharing practices, and high integration costs.

• Although there is also a rapidly growing demand for optimization across different data models, the legacy systems we have right now and the methods we currently use aren’t easily interoperable.

Current Developments

Recent advancements in ML are now opening the doors to more rapid and robust translation between different platforms. This approach promises to enable significantly optimized medical research projects and, of course, healthcare. But a lot more work needs to be done to achieve true interoperability across ML platforms.

Neural networks used in AI are essentially a set of algorithms that learn to recognize patterns. They leverage labeling, machine perception, or raw cluster inputs to interpret sensory data. Every time the algorithm recognizes a pattern, it essentially studies numerical data found in vectors where all real-world images, time series, sounds, and text are translated.

As such, it’s vital to develop distributed AI architectures with edge ecosystems and multi-cloud infrastructure offered by different service providers. Interoperability between startups and tech giants is also essential to ensure transparency and flexibility in the absence of appropriate regulation.

The ONNX community developed several different tools to convert and run different data models. You can quickly convert ML models that are trained on diverse frameworks to the ONNX format with tools like ONNXMLTools and make it work.

However, while the ONNX format accelerated the unification of AI and ML efforts of corporate giants, it quickly became apparent that having all this data in one place didn’t create a path to effortless success.

Semantic Interoperability

The future of ML will probably depend on semantic interoperability. Semantic interoperability describes the process of computer systems that exchange information with unambiguous, shared meaning. As such, high-quality data sets are vital to properly train ML models regardless of whether the data was aggregated from heterogeneous sources or a single source.

It’s important because algorithms can’t easily learn the patterns, anomalies, or make predictions when they are working with a combination of different data sources and when the information doesn’t exactly mean the same thing.

Conclusion

In order to provide coordinated and efficient data transmission between information systems, interoperability is essential. As you may expect, companies in the majority of industries can profit from this essential feature.

Without interoperability, the development of AI would be constrained and available to just big tech. This is due to the fact that only the biggest tech companies have access to the resources and, more crucially, the data that allows for ongoing and effective learning.

It is safe to say that the future of machine learning depends on semantic interoperability and global data machine communication standards.