NVIDIA Inception and Myukura’s R&D line: specialised models and local artificial intelligence

When people talk about artificial intelligence applied to healthcare, much of the conversation revolves around models.

In practice, building a product requires settling many other decisions: where each process runs, how long it takes, what data it needs, what information must leave the device, how the result is evaluated and what happens when the system gets it wrong.

Those are architecture questions, not only artificial intelligence ones.

Within Myukura, part of that work is concentrated in a line that combines specialised models, local processing and GPU-accelerated computing.

Our membership of NVIDIA Inception sits precisely there.

What NVIDIA Inception is

NVIDIA Inception is NVIDIA’s programme for startups developing products based on artificial intelligence and other accelerated computing technologies.

According to NVIDIA, the programme offers access to technical resources, training through the Deep Learning Institute, tools and SDKs, certain advantages on products and services in its ecosystem and opportunities to connect with its network.

The programme requires neither equity nor an entry fee.

For Myukura, however, one of its most concrete effects came from outside a platform or a technology credit.

It was through NVIDIA Inception that we learned about the call for the 8th MareNostrum Hackathon at the Barcelona Supercomputing Center.

From the programme to the MareNostrum Hackathon

Our selection for the hackathon allowed us to work for several days on high-performance computing infrastructure and to count on technical support from the participating mentors.

Our team’s goal was to study a specific problem: automatic speech recognition in a clinical context and in Spanish.

During the hackathon we trained and evaluated our own models and compared different configurations to better understand the limitations that appear when working with clinical vocabulary and different ways of speaking.

This kind of work is especially valuable because it forces us to move away from the abstract perception that “the model works” and turn it into measurable questions:

Where does it go wrong?

With what kinds of words?

What happens with medications?

What happens when the accent changes?

What computational cost does improving the result carry?

Why we develop specialised models

Part of Myukura’s research focuses on developing and evaluating our own models for specific clinical tasks.

Our work addresses problems such as speech recognition in clinical Spanish, the identification of medical terminology, medications, doses and other expressions especially sensitive to transcription errors, as well as the variability introduced by different accents and natural ways of speaking.

But a model’s size is not, in itself, a measure of its usefulness.

For certain tasks, we investigate specialised architectures able to offer greater control over the system’s behaviour, its computational requirements and the way it can be evaluated.

This becomes especially important when the goal is to run artificial intelligence close to where the data is generated.

A specialised, optimised model can reduce memory and computing requirements, improve latency and make deployment on local infrastructure easier.

Myukura’s architecture starts precisely from that idea: there is no single model suited to every problem.

Different tasks may require different models, sizes and execution strategies. Our job is to investigate that combination and determine which architecture offers the right balance between accuracy, efficiency, privacy and capacity for oversight.

The other line: local artificial intelligence

A second area of work is Myukura Edge.

The goal of this line is to investigate architectures in which certain processes can run on infrastructure located within the organisation’s own environment.

In healthcare, where processing takes place is not a purely technical matter.

It can affect latency, dependence on connectivity, control over the infrastructure and the path certain data takes.

That is why we study an architecture that does not necessarily depend on a single form of computing.

The cloud and local processing can have different functions within the same system.

An architecture before a model

The experience of these months has reinforced an idea we consider central to Myukura.

Clinical artificial intelligence should not be approached as a race to adopt the latest fashionable model.

To build a useful system you have to decide where each technology adds value and where it introduces new risks or complexity.

Our work with NVIDIA, access to GPU infrastructure and the MareNostrum Hackathon experience are part of that learning.

We are not looking for technology to replace the professional’s judgement.

We are looking to build tools that can help reduce administrative work while keeping the professional within the review process.

That combination of specialisation, architecture and human oversight will remain one of the foundations of Myukura’s R&D.

NVIDIA, the NVIDIA logo and NVIDIA Inception are trademarks or registered trademarks of NVIDIA Corporation in the United States and other countries.