Turbo-charging AI: Collaboration with NVIDIA renewed as joint tech center marks three years

'Finland is a great place to collaborate, the people are open and optimistic,' says Frédéric Parienté. 'Successful collaborations require funding, good researchers and hardware, of course, but also technology enablement, and that is what the NVIDIA AI Technology Center brings about.' Parienté, deputy director of NVAITC in the EMEA region, commented on the ease and productivity of the NVAITC Finland collaboration on a recent visit to the Aalto University campus, where NVIDIA, FCAI (Finnish Center for Artificial Intelligence) and CSC IT Center for Science met to renew their partnership.

Since June 2020, projects under the NVAITC umbrella have been diverse, from computer vision and differential privacy to natural language processing and Gaussian processes. The common thread is acceleration of computing with graphics processing units (GPUs) for specific scientific use cases.

As University of Helsinki and FCAI professor Keijo Heljanko commented following NVAITC Finland’s first year, 'GPU computing is a crucial enabling technology for AI, as the performance and energy efficiency of modern GPU architectures has made them invaluable tools for AI research. Collaborating with NVIDIA as part of NVAITC has given our researchers access to world class expertise in GPU computing, enabling new and optimized GPU-based AI algorithms to be developed in FCAI research projects.'

One major output is the Python package for differentially private probabilistic programming and its follow-up, cryptographically secure random number generation. The implications for speed and handling of sensitive data are significant, says NVIDIA’s Niki Loppi, one of the co-authors. 

Another advance concerns the Found in Translation project, where the researchers are tackling the scale-up challenge of multilingual translation models by introducing language-specific components to modern shared neural network architectures.

With FCAI and Aalto University assistant professor Arno Solin, NVAITC investigated spatial-temporal variational Gaussian processes. The result was a significant reduction in computation through parallelization and more efficient use of the underlying hardware. 'Working with NVAITC helped us push the research even further. With the help of Niki from NVAITC, we managed to get a thirty-fold boost to the computation speed of our methods, really squeezing out the full potential of the hardware. This also means that the reference implementation became more energy-efficient,' says Solin.

Juho Kannala, assistant professor in computer vision at Aalto, led a project to improve real-time depth perception. 'NVAITC helped us get our models running on mobile hardware that is commonly used in autonomous vehicles. They helped with adding missing features to the hardware libraries that we relied on. This enabled us to add a real-time demo of the methods in the research paper,' explains Kannala.

Beyond the 15 joint research projects conducted so far through NVAITC Finland, Parienté and Loppi highlight the education and training aspect of the collaboration, for example the “AI Meets X” webinar series and the master’s level project course at Aalto University that used the Modulus framework. 

NVAITC in Finland is clearly spearheading the melding of high-performance computing with AI, says Kimmo Koski, managing director of CSC, citing the fact that all the NVAITC projects have run on CSC computing resources. Computing power will continue to be crucial as dataset and model sizes dramatically increase, fueling the new revolution of large language models and generative AI. For these applications, the accelerated computing expertise of NVIDIA can be invaluable, says Parienté.

Projects in Finland have been high-quality and ambitious, says Loppi, while also emphasizing that NVAITC aims to work with academics and projects across a wide spectrum, event those who are just starting out using GPUs. To get in touch or submit a project, visit https://fcai.fi/nvaitc 


This article was originally published on FCAI's website.