Real-Time Machine Learning | 19 566

Real-Time Machine Learning (RTML) Grant Opportunity: FAQs

1) What is the Real-Time Machine Learning (RTML) grant opportunity?

The Real-Time Machine Learning (RTML) grant opportunity is a research solicitation focused on pushing machine learning beyond mostly offline, batch training workflows and into systems that can learn continuously from streaming data. The aim is to enable machine learning systems that can interpret data as it arrives, update models immediately, and use updated knowledge to handle unfamiliar situations, all while operating under tight time and energy constraints.

2) What core problem is the program trying to solve?

The solicitation targets a long-standing challenge in computing: building machines that can process incoming information in real time, learn from it immediately (not later), remain reliable as conditions change, and do so with far better energy efficiency than typical modern compute platforms. The program frames this as moving closer to brain-like energy efficiency rather than power-hungry approaches common today.

3) Why is "real-time" learning emphasized instead of traditional training workflows?

Because many next-generation applications cannot wait for periodic retraining cycles. In the RTML framing, data arrives continuously, conditions can change quickly, and decisions often must be made on the spot. Offline retraining and delayed model updates are presented as too slow or operationally impractical for these environments.

4) What kinds of applications does RTML consider important?

The opportunity highlights several domains where continuous, on-the-spot learning matters, including autonomous vehicles, defense and military systems, healthcare informatics, and business analytics. These are examples of settings where streaming data and rapid changes make delayed retraining unacceptable.

5) Is this program focused only on software or algorithms?

No. A key emphasis is that real-time learning is not just a software issue. The solicitation explicitly calls for new thinking in both machine-learning methods and the underlying hardware, arguing that conventional approaches can be too slow, too power intensive, or too dependent on periodic retraining for emerging streaming-data domains.

6) What does the solicitation mean by "co-design"?

Co-design means proposals are expected to treat algorithms and hardware as a coupled system rather than optimizing each independently. The technical focus is on developing algorithm-and-hardware pairs (or tightly integrated approaches) that enable continuous real-time processing and adaptation under real operational constraints.

7) What stages of the machine learning lifecycle does RTML want to happen in real time?

The program seeks foundations for approaches where all stages of training can occur in real time, not only inference. Specifically, it calls out:

• Incremental training: updating models as new data arrives
• Hyperparameter estimation: tuning model settings that in traditional pipelines often require expensive search
• Deployment: moving updated models into operation without long delays

8) How does RTML distinguish itself from recent advances in large-scale machine learning?

The solicitation notes that while recent years have produced impressive results and specialized electronic hardware for heavy computation (for example in speech recognition and computer vision), emerging RTML target domains look different. They involve continuously streaming datasets and operational constraints that can make conventional approaches too slow, too power intensive, or too reliant on periodic retraining.

9) What outcomes does the program seem to be aiming for?

Based on the description provided, ideal outcomes include architectures and learning methods that:

• keep learning after deployment
• stay stable and reliable as conditions drift over time
• maintain strong performance without massive energy use
• avoid requiring cloud-scale resources to remain effective

10) Who is sponsoring or running the RTML opportunity?

This opportunity is described as a joint effort between the National Science Foundation (NSF) and the Defense Advanced Research Projects Agency (DARPA). It is structured to encourage high-performance and energy-efficient solutions and highlights potential post-award collaboration between DARPA-supported and NSF-supported researchers.

11) What is the significance of NSF and DARPA being involved together?

The solicitation signals an intent to connect fundamental academic research (commonly associated with NSF) with mission-driven and hardware-forward development perspectives (commonly associated with DARPA), particularly for systems expected to operate in demanding real-world environments.

12) What is the funding opportunity number for this solicitation?

The NSF Funding Opportunity Number listed for RTML is 19-566.

13) When was this funding opportunity created, and what was the original closing date?

The listing states it was created on March 7, 2019, and the original closing date was June 6, 2019.

14) What type of funding opportunity is RTML described as?

It is described as a discretionary grant solicitation from NSF.

15) What is the activity category for this opportunity?

The activity category is listed as science and technology and other research and development.

16) What CFDA numbers are associated with RTML?

The CFDA numbers listed are 47.041 and 47.070.

17) How many awards were expected?

The opportunity summary indicated an expectation of around 12 awards.

18) Is there an award ceiling specified in the provided summary?

No. The provided summary field indicated that an award ceiling was not specified.

19) Who is eligible to apply based on the provided information?

Eligibility is listed broadly as "Others", with additional clarification referenced in the full solicitation text. This implies applicants would need to consult the official solicitation for exact eligible organization types and any constraints.

20) What makes "energy efficiency" a central theme in RTML?

The program emphasizes that real-time learning must operate under tight energy limits and seeks approaches that are closer to human-brain-like efficiency than to modern power-hungry compute platforms. This focus is tied to practical deployment needs in real-world environments where power and thermal budgets can be constrained.

21) Does RTML focus on inference-only acceleration?

No. The solicitation is explicit that the goal is not just improved inference. It seeks next-generation approaches where training itself (including incremental updates and hyperparameter estimation) can occur in real time alongside deployment.

22) What does the opportunity suggest about operating conditions in target environments?

It highlights that conditions can change quickly, data arrives nonstop, and decisions may need to be made immediately. It also indicates an interest in systems that remain stable and reliable as conditions drift, implying robustness over time is an important consideration.

23) Does the solicitation mention any collaboration expectations after awards are made?

Yes. It highlights the potential for post-award collaboration between DARPA-supported and NSF-supported researchers, suggesting an emphasis on cross-community interaction between fundamental research and mission-driven development perspectives.

24) In one sentence, what is RTML essentially calling for?

It is essentially a call for research that makes real-time, continuously learning machine intelligence practical by inventing new algorithm-and-hardware pairs that can train, adapt, and deploy on the fly under real operational constraints.