U.S. policymakers have begun looking at ways to reinvigorate our national approach to research and innovation, including harnessing the incredible potential of AI. But our nation’s AI capabilities are only as good as the foundation they’re built on — computing power and data. And the enormous cost and complexity of maintaining those two ingredients has made access to them extremely limited. This means that novel research insights can increasingly only be unlocked by well-funded incumbents.
What was once a promising accelerator for research breakthroughs has been tempered by market concentration and vendor lock-in, which has increased costs and constrained access to tools that underpin basic AI research. But without broad-based foundational research carried out in partnership between government, industry, and academia, many of the inventions that we rely on today – such as GPS – might have been left on the cutting room floor.
Democratizing access to a shared computing infrastructure would ensure that future AI advances happen in the U.S. To address this issue of access, Congress passed the National AI Research Resource (NAIRR) Task Force Act of 2020, which mandated the creation of a task force with the remit of delivering to Congress a plan for how to democratize access to computing power and data – and ultimately ignite AI R&D.
The NAIRR is about bringing together data, resources, people, and expertise to advance leading-edge AI research in the U.S. If the NAIRR succeeds in its task of broadening affordable access to computational power and large government datasets, it will bolster basic scientific research, broaden AI innovation to groups and institutions that have yet to benefit, and it will allow the U.S. to compete in the global race to develop emerging technologies.
In July, IBM responded to a request for information (RFI) from the National Science Foundation (NSF) and the White House Office of Science and Technology Policy (OSTP) on initial findings and recommendations of the NAIRR task force. First and foremost, our response includes technical recommendations for how the NAIRR can enshrine open and cost-effective principles to maximize access to data, compute capacity, and software. But IBM goes one step further to incorporate lessons gleaned from the COVID-19 High-Performance Computing Consortium, and the National Strategic Computing Reserve Blueprint, to focus on the human capabilities – such as user training and access – that will ultimately break down barriers to AI R&D to supercharge our AI innovation ecosystem.
As the work of the NAIRR task force builds to the release of its final report to Congress in December, IBM offers four suggestions for the NAIRR:
1. Make It Open: To mitigate startup costs and scale quickly, the NAIRR should be built using commercially available cloud offerings, expanding on proven models such as CloudBank and STRIDES. Due to its massive scale and need to handle large government datasets, the NAIRR must seamlessly combine and orchestrate multiple public and private clouds, and on-and off-premise infrastructure. This can only be accomplished by adopting an open hybrid- and multi-cloud architecture, and a federated access model. An open hybrid cloud architecture would allow multiple existing public clouds to be integrated into the NAIRR, and it would prevent any one provider from establishing a monopoly over the resource over time. And by adopting a federated access model, the NAIRR would create a uniform environment for compute providers to provision and share services. Federation best integrates the diverse resources of the NAIRR, decreases costs to taxpayers, and fosters collaboration among researchers, program administrators, and service providers alike.
2. Ensure Flexibility to Grow and Iterate: If the NAIRR is to maximize its benefit to researchers and institutions, it will need to obtain buy-in from diverse stakeholders inside and outside of government. And it must do so over an extended time horizon. The NAIRR must be built as a flexible and iterative resource capable of integrating new technologies, tools, and services from the rapidly growing field of AI research. Adopting a federated approach to implementation, deployment, and administration would incentivize computing providers to offer the latest tools, software, and resources – and keep costs in line.
3. Build Human Capital: Building and funding the physical infrastructure of the NAIRR will mean little if AI researchers are unable to use it. In addition to physical infrastructure, the NAIRR must build a reservoir of educational tools and services, user support communities, and easily accessible user access mechanisms to encourage its use and mainstreaming. In Its RFI response, IBM proposes the creation of a Research Resource Skills Academy, along the lines of the IBM Quantum Education and Research Initiative, which partners will 12 historically black colleges and universities to develop education, community resources, and technical communities to power the field of quantum computing.
4. Democratize Access: A core tenant of the NAIRR is ‘democratizing’ access to AI R&D. But achieving widespread use and true democratization of the NAIRR will only be achieved if the resource is stood up as an open, federated, and iterative resource. Ultimately, the NAIRR will not be only ‘democratize’ access to a research resource, it could also show how AI resources are being used, unearth new research questions, and be a transformative catalyst for investment that bring together academic, government, and industry participants.
Read IBM’s full response to the National Artificial Intelligence Research Resource RFI here.
-Jeffrey Brown, Science & Technology Policy Executive, IBM
-Talia Gershon, Director, Cloud Infrastructure Research, IBM Research
Share this post: