Three senior defense officials provided the latest update on that nascent effort to a small group of reporters during a media roundtable at the Pentagon on Monday.

Details they shared suggest the Army is evolving its approach to known and unknown dangers associated with deploying AI, via Project Linchpin. And in parallel, officials are also producing a new “AI risk reduction framework” to inform all future pursuits.

“[This is] in line with a lot of the work that the White House has pushed out with AI, and that the DOD has pushed out about responsible AI with Task Force Lima and all those types of things. We’re definitely embedded with all those things, but we’re also looking at what are the second- and third-order impacts of things that we’re going to have to address earlier, from an obstacle standpoint,” Young Bang, principal deputy assistant secretary of the Army for acquisition, logistics and technology, explained.

First conceptualized in 2022, Linchpin is ultimately aimed at generating a safe mechanism to continuously integrate government- and industry-made AI and machine learning capabilities into Army programs. 

“Think of Project Linchpin as our path to delivering trusted AI,” Bharat Patel, product lead for Project Linchpin at the Army’s program executive office for intelligence, electronic warfare and sensors, told reporters.

“If I can leave you with something right up front — it’s literally all the boring parts of AI. It’s your infrastructure, it’s your standards, it’s your governance, it’s your process. All of those areas are things that we’re taking on, because that’s how you can tap into the AI ecosystem and that’s how you deliver capabilities at scale,” Patel said.

The Army’s Tactical Intelligence Targeting Access Node (TITAN) program, which encompasses its next-generation ground system to capture and dispense sensor data for sensor-to-shooter kill chains, marks the first program that officials seek to enable with algorithms affiliated with Project Linchpin.

“I would say we’re, right now, just collecting AI use cases. TITAN is expected to support a certain theater. We’re working with that theater and that program to determine everything — kind of the left and right limits, and how would we deploy — all that is happening now. But if you think about it, for classic computer vision problems, each theater is different. You can’t think a model for [European Command] is going to work out of the box for [Indo-Pacific Command]. The trees are different, the biosphere is different, all that is different. That’s why it’s super important to get after the use case and where that [area of responsibility] is specifically at. So, we are looking at that very closely because we want to make sure we tailor the model to support the customer,” Patel told DefenseScoop during the roundtable.

Bang, Patel and their team have been conducting what they called “a ton of market research” as part of standing up this new program. Since Nov. 2022, they’ve released four requests for information on Project Linchpin, collected “well over” 500 data points, and met individually with more than 250 companies. 

Momentum will continue to build in those aspects in the near term — and possibly also through budget bumps, according to Matt Willis, the director of Army prize competitions and the small business innovation research (SBIR) program.

“In [fiscal 2025], in the next year or so, we’re predicting a significant investment in our SBIR program towards AI in particular — again, strategically aligned with Project Linchpin, [that’s] potentially up to or more than $150 million. So, that’s about 40% of the program, and this really demonstrates our commitment to innovation, to AI and how small businesses across the country can certainly contribute to the Army,” he said.

At the roundtable, the officials also repeatedly emphasized their intent to confront ethical and security risks associated with AI and machine learning with Project Linchpin as it continues to mature. 

In that sense, Army officials are also crafting an “AI risk reduction framework” that Bang noted will be designed to get at Army-specific “obstacles” that accompany deploying the emerging technology.

“It’s really a way to identify the risks and mitigate some of those risks — to include data poisoning, injections, and adversarial text attacks. Now, specifically, are you asking ‘Have there been those types of things that we found in Linchpin?’ There are those types of things that we know are out there in the environment or the enterprise. And so whether it’s commercial or on the DOD side, we know they’re out there. So we’re actually trying to mitigate some of those,” Bang told DefenseScoop.

“It’s really a framework to look at what are the cyber risks and vulnerabilities associated with third-party algorithms, and how do we work with industry to categorize that to look at tools and processes to reduce the risks, so then now we can adopt that faster?” he added.

His team has also been hosting a number of engagements with their industry partners to figure out a path forward with a potential need to request AI bill of materials, or AI BOMs from companies.

Such resources are envisioned to essentially help the government better understand potential risks or threat vectors the capabilities could introduce to their networks.

“We’re, again, conducting more sessions with industry. We understand their perspective. And it’s not to reverse engineer any [intellectual property], it’s really for us to get a better handle on the security risk associated with the algorithms. But we do understand industry’s feedback, so we are working really more on an AI summary card. You can think about it more like a baseball card. It’s got certain stats about the algorithm, intended usage and those types of things. So it’s not as detailed or necessarily threatening to industry about IP,” Bang said.

The post Army rethinks its approach to AI-enabled risks via Project Linchpin appeared first on DefenseScoop.

The service’s Project Linchpin aims to provide an artificial intelligence and machine learning development and delivery operational pipeline for intelligence, cyber and electronic warfare sensor systems.

An RFI posted Monday on Sam.gov seeks information about computer vision technology.

“The Army’s pivot to Large-Scale Combat Operations … has underscored a critical need for improved capability to identify, monitor, and strike targets from farther distances with enhanced precision and with sharply reduced sensor-to-shooter timelines. Deep sensing is a crucial enabler for this capability. The Army also requires enhanced Line-of-Sight sensor to shooter capabilities. To address these capability needs, Project Linchpin is requesting information on object detection and computer vision capabilities to support Army initiatives,” according to the notice.

That includes capabilities to detect and classify objects using labeled datasets, using imagery from horizontal full-motion video, electro-optical and infrared sensors mounted on ground combat vehicles, as well as overhead electro-optical, infrared and synthetic aperture radar imagery captured by satellites.

The RFI is intended to help “scope” a future contracting activity to train and provide artificial intelligence and machine learning models for Army programs, according to the notice.

Meanwhile, another RFI posted Monday on Sam.gov aims to get feedback from industry to inform ideas for implementing and automating what the Army is calling an artificial intelligence bill of materials, or AIBOM, that would address potential vulnerabilities in the “supply chain of components” that are used to build AI models.

“The AIBOM is a novel concept created by Project Linchpin and will require feedback from industry to mature,” the notice states.

Young Bang, principal deputy assistant secretary of the Army for acquisition, logistics and technology, shed light on the service’s thinking during a meeting with DefenseScoop and other reporters at the Army’s Technical Exchange Meeting in May.

“We’re toying with the notion of an AIBOM [program]. That’s because really, we’re looking at things from a risk perspective. Just like we’re securing our supply chain, semiconductors, components, subcomponents, we’re also thinking about that from a digital perspective,” Young said.

According to the RFI, the envisioned AI bill of materials would be informed by a software bill of materials that includes details about supply chain relationships of the components used to build and validate” a model; information about the model’s properties, architecture, training data, hyperparameters and intended usage; and the “lineage and pedigree of the data” used to make the model.

“The intent is to say, ‘can we look at the observability, the traceability of how you actually develop the algorithms, what are the features and the parameters you tested, what are the datasets that you use to ensure we have more trusted … outcomes, and that there’s no risk like Trojan triggers, poisoned datasets, or prompting of unintentional outcomes over the algorithm. We really need to think about that,” Bang said at the technical exchange meeting.

The Army is asking industry to weigh in on the pros and cons of the proposed AIBOM concept; technical means to implement and automate the components of the bill of materials; tools, processes and skills required to put it into an AI and machine learning operational pipeline; and a cost estimate for what it would take to produce an AIBOM.

The response deadline for both RFIs is Dec. 1.

The release of the special notices comes just a few weeks after the Army announced that it had awarded the first contracts for Linchpin to Booz Allen Hamilton and Red Hat to support research and development related to the principles of “traceability, observability, replaceability and consumability,” aimed at promoting data integrity and a modular open system architecture.

The awards marked “a significant first step to decouple AI from software, decompose components within a MLOPs pipeline, and wrap layers of security around the entire process. These design principles will allow the Army to leverage the best of breed technology available across industry, academia, and government” Col. Chris Anderson, project manager for intelligence systems and analytics in Program Executive Office Intelligence, Electronic Warfare and Sensors, said in a statement Oct. 2.

The post Army issues new RFIs for Project Linchpin artificial intelligence initiative appeared first on DefenseScoop.

Booz Allen Hamilton and Red Hat both won a contract through a Broad Agency Announcement worth a combined value of $2 million to support research and development for Project Linchpin, the Army said in a release Wednesday.

Project Linchpin aims to provide an artificial intelligence operations pipeline for Army programs, beginning with the Tactical Intelligence Targeting Access Node (TITAN) program.

The deal has a six-month period of performance along with an option of no more than five years. Specifically, the contract deals with research for principles of Traceability, Observability, Replaceability and Consumability (TORC), according to a release from Program Executive Office Intelligence, Electronic Warfare and Sensors.

That framework aims to ensure model and data integrity, data openness as well as modular open system architecture design, the Army said.

“PEO IEW&S will use Project Linchpin to develop and deploy trusted AI & ML capabilities to intelligence, cyber, and electronic warfare sensor systems. This contract is a significant first step to decouple AI from software, decompose components within a MLOPs pipeline, and wrap layers of security around the entire process. These design principles will allow the Army to leverage the best of breed technology available across industry, academia, and government” Col. Chris Anderson, project manager for intelligence systems and analytics, said in a release.

The Army has been conducting a variety of industry engagements and releasing requests for information over the last year or so in an effort to better understand what exists in the commercial world to inform Project Linchpin.

The post Army awards contracts for first AI program of record appeared first on DefenseScoop.

This is where the Army’s Project Linchpin comes in.

“In the simplest form and concept it’s the way how we are going to deliver AI to the sensor modernization programs,” Bharat Patel, product lead for Project Linchpin at program executive office for intelligence, electronic warfare and sensors, told DefenseScoop in an interview.

Project Linchpin is the Army’s first program of record providing an artificial intelligence operations pipeline. The Army did not respond to multiple requests for information about how many AI programs of record it now has.

“[T]his project takes the standard AI and machine learning operations pipeline from the technology industry and modifies it to perform in a secure government environment while protecting operational data. Project Linchpin offers a secure structure that could be replicated across the Army to deliver AI at scale,” Patel and Maj. Nick Bono, an acquisition officer and systems coordinator for intelligence systems, wrote in a recent article posted to War On The Rocks.

Patel explained to DefenseScoop that while there is nothing novel about an artificial intelligence or machine learning ops pipeline in the commercial world, bringing the concept to government and the military in particular is the challenge.

“Industry has been doing this, we’re going to apply industry best practices, we’re going to leverage commercial technologies to establish this,” he said. “I’m going to decouple all the components within the pipeline and then compete that against industry best of breed, and then bring in industry technology developers to develop models against our sensor modernization programs.”

Patel and Bono wrote that the end state is for Project Linchpin to be a standalone program of record providing a government-owned pipeline that will mitigate risks by breaking up the vertical integration of efforts that have be in place historically.

Project Linchpin will create a disciplined approach that enables continuous integration of machine learning capabilities into Army programs.

“The key really behind that is machine learning — true machine learning — is almost never done. It’s even more rapid than software. Machine learning, you could develop the model, you create the model in a machine learning pipeline, then you put that model in its intended end state,” Patel said. “You put that model in the intended end state, but once the operational data actually hits it, that’s when you see the model performing. That’s when you see what we call model drift — in certain cases, sometimes the model is not performing exactly because there’s a new type of data that comes in. So we have to take that data, bring it back into the operational pipeline, retrain the model and then get that model back into the environment.”

Ultimately, the project seeks to answer a simple question: “How does the Army deliver secure, trusted AI/ML capabilities to AI-enabled systems?” Patel said.

“Project Linchpin recognizes that there are limited resources (funding, technology like high-powered computing, and human capital in the form of AI experts and engineers) for every individual AI-enabled system to manage the process of developing, training, and deploying AI/ML capabilities on their own.  Project Linchpin scales that effort to consolidate AI talent and resources in a secure environment.  The end state is the reliable delivery of high-performing AI/ML capabilities that users can trust to employ on their systems,” according to Patel.

The effort can trace its roots to Project Maven, an Air Force-led initiative that was created in 2017 to accelerate the adoption of artificial intelligence. It focused initially on processing full-motion video feeds from unmanned aerial systems in the Middle East to support U.S. counterterrorism and counterinsurgency wars. Staff within the program office had been working with the Department of Defense on Project Maven since its inception, leading operational pilots and providing technical oversight.

After the work on Project Maven and coupled with the Army’s push for sensor modernization programs, officials began developing requirements that either implied or specified a need for machine learning, Patel said.

“Going back to my previous knowledge of actually working in this space, in the DOD, and how challenging it was in how complex the different types of tools, technologies [and] industry partners you need to bring in, I made the case to [the program manager] that it was pretty much unaffordable for each PM to go do their own machine learning pipelining,” Patel said.

As a result, he made the case to the PEO last August, briefed it to the principal deputy assistant secretary of the Army for acquisition, logistics and technology last September, and by October, it was announced that the service would be standing up Project Linchpin as a program of record.

Operational outcomes for JADC2

Officials hope Project Linchpin capabilities can ultimately be used across the entire Army.

The main effort to begin with is the Army’s Tactical Intelligence Targeting Access Node (TITAN) program — the service’s next-generation ground system to collect and disseminate sensor data to improve sensor-to-shooter kill chains. It will be a key aspect of not only the service’s so-called deep sensing priority — the ability to discover targets across thousands of miles for long-range fires — but the Pentagon’s new way of war known as Joint All-Domain Command and Control (JADC2).  

Patel noted that there’s many components to sensor-to-shooter networks, and an artificial intelligence operations pipeline can help optimize them.

“Our goal is owning that front part of sensor to the shooter is really optimized the ability to identify objects of interest and turn them into — I don’t want to call them targets, but targets of interest,” he said.

For example, an airborne sensor could be sending data to TITAN. As part of that process, TITAN must have algorithms to quickly identify an object and send that information to the shooter. But this capability can eventually be replicated in the aircraft, Patel said, forecasting the eventual goal of perpetuating this pipeline throughout the entire Army.

“Eventually, it’s taking that model and then bringing it into the aircraft. We want to do it for TITAN first and then bring it into the aircraft so that it’s identifying objects at a much more rapid pace,” he said.

Focusing on TITAN was in some ways a strategic decision given that it will receive data from space, air and ground sources. This means officials can build cross-cutting models that span several domains and program offices.

“One of the use cases that we brought to the table was, each PM does not need to build their own machine learning pipeline capability. In addition to that, each PM, even if we have a central pipeline, they don’t need to be in the model development business because this model that was developed for TITAN against [synthetic aperture radar] data, I could just optimize it to work for the aerial side,” Patel said. “This aerial sensor doesn’t need to go out and find its own company to build its own SAR model. I’m just going to tweak and optimize and put it on that system. We’re focusing on TITAN is the right place, because we’re going to have all bunch of sensor data and then we’re just going to optimize and provide all the AIs to the various programs.”

For the time being, Project Linchpin doesn’t appear as its own standalone entity in the Army’s fiscal 2024 budget documents. The only reference to it falls within the TITAN program.

Within TITAN, the documents note that the program will “initiate development and prototyping of Artificial Intelligence/Machine Learning (AI/ML) platforms (i.e., Project Linchpin),” and that a fiscal 2024 funding increase for a portion of the program represents the advancement of artificial intelligence and machine learning integration capabilities through the establishment of Project Linchpin.

Patel and Bono warned, however, that implementation of this pipeline will be critical.

“In most commercial applications the risks of error are just an inconvenience. If Google Maps recommends a faster route that takes you into a traffic jam you might simply be late for your meeting. However, if an intelligence sensor confuses a school bus for a tank — or fails to detect the tank at all — the results could be catastrophic,” they wrote. “No matter how much talk there is about how game-changing AI will be in the future, failing to plan for AI’s ‘back end’ infrastructure is planning to fail.”

In order to get the right data to build off, the program will be relying on the Army’s Office of the Assistant Secretary of the Army for Acquisition, Logistics and Technology (ASAALT) data mesh concept, which aims to essentially standardize the vast quantities of data the service has in order to make it easier to get it to the right decision makers at the right time.  

Until that is fleshed out, Patel said officials will be challenged from a data perspective.

“We’re going to be focused on data collection teams in the interim. These data collection teams are going to be responsible for working with each PM to go … grab some data and bring that into our repository,” he said.

To train the algorithms properly, they might only need a fraction of the hours and hours of full-motion video, for example, because that small tidbit is really important to the algorithms’ development. The rest isn’t. To do that would be extremely manual.

But when ASAALT’s model comes online, that should relive some of that burden, Patel said.

“We are going to be challenged, but we have a path to address some of that as it relates to the sensor part of it. But then once ASAALT comes online with the data fabric and data mesh, I think it’s going to start being a little bit cleaner,” he explained.

Non-traditional contractors

When it comes to industry support, Patel said officials are looking primarily at non-traditional contractors and small businesses.

“We’re constantly going out to industry looking for who has the best technology for model training, who has the best technologies for model development? That’s right for the non-traditionals and small businesses,” he said.

The most recent request for information went out to industry in March, and the team is planning to use an upcoming Technical Exchange Meeting in May to provide more information.

Moreover, the Project Linchpin team briefed the effort to Army senior leadership in April with additional discussions planned going forward.

The post Project Linchpin aims to set Army on sustainable path toward integrating AI into weapons programs appeared first on DefenseScoop.