The application of artificial intelligence techniques spans every conceivable field. With data usage at its core, the significance of its encoding is paramount, as elucidated by Leonardo Chiariglione, the scientist renowned as the "father" of the MP3 format

In the expansive realm of AI, data encoding emerges as a critical and valuable component. No industry remains untouched or unaffected by artificial intelligence methods. Data forms the cornerstone of AI, with vast amounts being generated annually, though not all are pertinent. It’s therefore crucial to encode and interpret these data effectively to maximize their utility. This necessity gave rise to MPAI (Moving Picture, Audio, and Data Coding by Artificial Intelligence), a non-profit international body dedicated to promoting efficient data utilization.

The architect and leader of this initiative is Leonardo Chiariglione, a distinguished engineer and scientist, globally celebrated for developing the MP3 audio compression algorithm with the MPEG group, which he established and led for over three decades. Reflecting on MPEG, Chiariglione notes:

“It’s the group that has crafted an impressive array of standards and technologies, birthing an industry valued at several hundred billion dollars.”

In 2018, he estimated the annual global value of products and services based on MPEG standards at 1.5 trillion dollars, roughly 2% of the world’s GDP.

Transitioning from MPEG, primarily focused on data compression, the shift to MPAI occurred three years ago, centering on data comprehension. According to MPAI, its mission involves developing standards for data encoding. Data must be digitally represented to facilitate easy access and extraction of necessary information. In today’s era of big data, this often means distilling few valuable bytes from terabytes of data.

Data Encoding with AI: The MPAI and AI for Health Initiatives

MPAI has launched 16 projects, with standards already established for nine and ongoing efforts for the others. One notable project is AI for Health (MPAI-AIH).

This initiative envisions a system where devices (clients) gather and process individual health data using shared AI models and upload this information to a backend, accompanied by licenses articulated through smart contracts. Third parties can then access and process this health data under the specified licenses. The backend periodically employs federated learning to collect AI models and retrain shared models, which are subsequently redistributed to all clients. Chiariglione elaborates:

«This system represents individuals, potentially patients, through their smartphones, equipped with the MPAI-AIH platform for data processing via AI techniques. This processing is modular, with established interfaces, interconnected and executed as workflows within an AI framework. Practically, a person acquires personal (health) data, which is processed: some is retained, while other data is sent to the backend under licenses defined by smart contracts, thus ensuring privacy».

Each participant receives models for their smartphone, consisting of self-learning neural networks. By engaging in regular activities, users train these models, which are then sent back to the backend, safeguarding sensitive data and maintaining user privacy.

«The advantages are manifold: the community’s efforts in training networks on their data are shared upon distribution of the new model. Users gain detailed insights, and the community benefits from shared, commonly useful data. Additionally, through federated learning, the redistributed model incorporates collective intelligence, allowing the system to continually learn and evolve».

For end-users, this translates to access to increasingly refined information and services. For a hospital leveraging this system, it means possessing processed data critical for developing, say, a clinical trial. Additionally, third parties stand to gain in terms of data access and shared knowledge.

Interviewer: Could you share the story behind MPAI’s inception, especially in relation to MPEG’s legacy?

Chiariglione: Reflecting on the origins of MPAI, I see it as both an evolution from MPEG and a fresh venture into data encoding standards powered by AI. Having founded MPEG in 1988 and led it for over three decades, I witnessed its transformation into an entity overly focused on patent royalties. MPAI was born out of a desire to return to MPEG’s original vision, prioritizing the creation of practical and fair standards, balancing intellectual property rights with broad accessibility.

Interviewer: Regarding the data encoding based on artificial intelligence, what potential does AI unleash in the fields where it is implemented?

Chiariglione: It’s vital to acknowledge that our focus is not limited to a singular field. Rather, we wield a technology—or more precisely, an ensemble of technologies—that we adapt and apply across a diverse array of domains, including but not limited to the metaverse, connected autonomous vehicles, portable avatars, and several others.

Interviewer: In relation to the metaverse, what areas have you concentrated on?

Chiariglione: The standard released in October paves the way for the creation of metaverses characterised by defined interoperability. A metaverse encompasses numerous processes, each detailed by the standard with functional specifications. Hence, for every process, we ensure it accomplishes certain tasks and generates specific data. The data’s format remains unspecified, due to the nascent nature of metaverse technology. Our role is to facilitate data production with distinctive characteristics that enable the generation of a unique environment. This is structured so that if another entity constructs an additional metaverse with data of similar functional traits – albeit through different technologies – interoperability becomes feasible via a data conversion service. Essentially, we facilitate functional interoperability among metaverses.

Technically, the architecture delineates terms and definitions, operational models, functional requirements of processes, actions, items, data types, and functional profiles facilitating the interoperability of multiple metaverse instances

Interviewer: Concerning the connected autonomous vehicle, on which aspects have you laboured?

Chiariglione: The standard we’ve crafted delineates the architecture of a connected and autonomous vehicle, marking the first in a sequence of planned standards. The objective of the MPAI-CAV (Connected Autonomous Vehicle) – Architecture standard is to specify the subsystems and components of such vehicles, thereby expediting the development of explainable components within the industry.

This standard identifies four subsystems, including one where the passenger interacts with the machine and another where the vehicle constructs its environmental representation. This representation mirrors a metaverse, accessible to other vehicles to enhance environmental comprehension. The final subsystem executes instructions from the third subsystem, effectively serving as the vehicle’s brain.

Today’s automobile industry primarily focuses on component integration. Vehicle manufacturers predominantly act as component integrators. This standard aims to define basic component interfaces, enabling developers to optimize and manufacturers to integrate these components.

In essence, we have “deconstructed the automobile,” establishing functional specifications for parts within a complete system, thus laying the groundwork for integration.

Interviewer: Regarding data encoding based on artificial intelligence, another standard you have developed concerns neural network watermarking. Could you elaborate on this?

Chiariglione: Neural network watermarking is a technological advancement that allows alteration of a neural network—for instance, to verify if a particular network has been duplicated. Current estimates suggest that custom AI solution development can range from 6,000 to 300,000 dollars. Therefore, it becomes crucial for owners to ensure traceability and integrity of neural network usage.

The ability to embed an invisible watermark, which does not degrade the network’s quality, is a nuanced requirement. Moreover, another fascinating aspect of the watermark is its capacity to imprint on outputs processed through a neural network. For example, with ChatGPT, it’s feasible to determine whether a text, even a half-page long, was generated by this specific Large Language Model or another. The MPAI Neural Network Watermarking standard facilitates measuring the watermark’s payload size in terms of the neural network’s performance impact, change resistance, and the processing cost of watermark injection.

Interviewer: In a prior interview, you referenced the immense value of MPEG standards-based services. What potential value do products based on MPAI standards hold?

Chiariglione: Discussing this matter is somewhat premature at this juncture. It’s crucial to remember that our inception occurred merely three years ago, compared to MPEG standards, which were the culmination of three decades of evolution. Presently, I can assert that MPEG primarily addressed a specific sector: media. However, we now span nearly every sector, thanks to the pervasive capabilities of artificial intelligence techniques.

Whereas the economy previously revolved around specific activities, today’s landscape is predominantly data-driven, underscoring the significance of interpreting and shaping this data. This is where artificial intelligence becomes a pivotal player.

Interviewer:  You are universally recognized as the “father” of the MP3, which dates back approximately 30 years. Currently, you are involved in the standardization of data encoding based on artificial intelligence. How do you envisage the near future?

Chiariglione: It’s challenging to predict. Just consider the developments in the past couple of years with the emergence of Large Language Models. We are witnessing genuine, unpredictably emergent discoveries. Hence, it is complex to speculate on what will transpire from the long-term interaction between human intelligence and data

Interviewer:  How do you perceive the evolution of standardization in data encoding based on artificial intelligence?

Chiariglione: The significance of standardization within the realm of artificial intelligence diverges considerably from its historical role. The complexity today far surpasses that of the MPEG era. A predominant portion of MPAI’s efforts is devoted to defining component interfaces, facilitating their integration into more elaborate systems, and thereby creating progressively intelligent machines.

Interviewer: So, can we assert that the AI-based standard is advancing towards what might be termed the… fourth dimension?

Chiariglione: Indeed, that seems to be the case. In the realm of artificial intelligence, we are navigating a much more expansive space, which can aptly be described as the fourth dimension.