United States

The Navy is working to streamline ship maintenance to achieve better fleet readiness and a “north star” of 75 mission-ready ships. This requires advanced analytics which leverage large maintenance datasets for predictive, proactive and preventative decisions. However, existing maintenance processes produce low quality data that prevents their use and limits insights from analytics. Efforts to improve this data through post-hoc curation have proven costly and time-consuming while failing to address the key problem: recording high-quality data in the first place. RUBICON (Rubric-Based Intelligent Curation) addresses this key problem by providing real-time AI-assistance and feedback that helps sailors adhere to data requirements as they record maintenance reports. Our prototype system focuses on 2-Kilo Maintenance Action reports (2Ks), which are written for every maintenance action taken on a ship and currently suffer from ubiquitous imprecise language, technical errors, and inconsistencies. RUBICON uses multiple agents and a large language model (LLM) to format, standardize, and organize text according to downstream data requirements and SME-informed best practices. This allows sailors of wide-ranging experience levels to produce expert-quality reports while entering simpler, unstructured descriptions of maintenance issues. As a result, sailors using RUBICON spend less time on report writing, and the reports they create support deeper insight. 

By improving the quality of data-rich 2Ks, RUBICON unlocks better analytics that will empower proactive decision-making. For example, improved equipment reliability estimates can empower Navy Program Executive Offices (PEOs) to choose contracts that reduce long-term labor and cost. There are several challenges for building our system in the 2K domain: (1) reports use language, terminology, and knowledge that is out-of-distribution for most LLMs; (2) existing data quality documentation does not adequately cover the diverse 2K error-types and report topics; (3) fast report-writing is critical for sailor relevance; and (4) sailors have diverse experience levels and needs. To address these challenges, RUBICON combines several innovative technologies and approaches: (1) the user interface converts LLM interactions into adaptive hints and user-prompts while avoiding slow chat-style workflows; (2) report quality rubrics (RQRs) encode requirements as natural language statements directly pulled from SME conversations; (3) adaptive retrieval systems uses structured and unstructured methods to standardize language and provide context to LLMs; and (4) separate information extraction and report generation steps reduce hallucinations in drafted reports. Currently, we are testing our prototype through live sailor demos for in-depth feedback. In this talk we will discuss the key features of RUBICON, describe the underlying technologies, and provide a demonstration of the user-interface.

NAML 2026

Rubric-Based Intelligent Curation (RUBICON): Real-time sailor assistance for high-quality maintenance report writing

2-kilo reports

human-ai teaming

data curation

maintenance

large language models

### Welcome to the 10th annual workshop on Naval Applications of Machine Learning!

NAML is the premier annual event showcasing current machine learning research relevant to Naval applications. We welcome participation from the defense community, industry, and academia. The NAML workshop is an opportunity for researchers from Navy research and development centers and our partner organizations to exchange technical ideas and to come together as a team. 
Naval Information Warfare Center Pacific is partnering with AFCEA San Diego chapter to host the workshop.
#### On-Demand Access

* This is an on-demand site and all the recorded content from the workshop is available here. 
* Access to the recordings is exclusively available to registered participants.
* Each registrant has received an email notification, with instructions on how to access the content.

[![](https://assets.underline.io/markdown_image/1/image/60cd12aa376c191d3467ea2ee60944ea.png)](https://naml2026.org/)

The goals of NAML are:

* To spread awareness of current machine learning research relevant to Naval applications
* To connect machine learning researchers with experts in military requirements
* To build and strengthen collaborations within the defense research community

#### *"What we do as a Navy is a human business. Don’t pass up an opportunity to come together as a team. Don’t pass up leadership opportunities, and don’t outsource them to a PowerPoint. If you have an opportunity to come together as a team, to increase belonging, do it."*
Admiral John M. Richardson, 31st Chief of Naval Operations
<br><br>

This is an on-demand site. All recorded content from NAML 2026 is available here and access to it is exclusively available to registered participants.

**Please log in using the email address you registered with.** If you have any issues logging in, please contact us at <naml2026@underline.io>.

Let’s celebrate the past decade by investing in the next. 🚀


Please log in to explore this event.

This is an on-demand site. All recorded content from NAML 2026 is available here and access to it is exclusively available to registered participants.

**Explore the latest advancements in machine learning tailored to Naval applications 🔎**

With on-demand access to NAML 2026, you can watch all lectures & browse poster presentations at your convenience, while also engaging directly with presenters via the Underline platform - ask questions, exchange ideas, and stay connected.

**If you still haven’t registered, you can do so now.**
👇
**Take advantage of a special one-time offer: full access for just $99, with complimentary registration for students, government, and military employees.**

**Get access here: https://naml2026-ondemand.paperform.co/**

Please check your spam folder, or contact our support team at **<naml2026@underline.io>** if you registered, but still haven't received an access email.

Let’s celebrate the past decade by investing in the next. 🚀

Please register!

The 10th Annual Workshop on Naval Applications of Machine Learning (NAML 2026) was held in San Diego, California, March 2-5, 2026. NAML is a forum for interdisciplinary experts from government, industry, and academia to discuss results of ongoing research on theory and applications of machine learning to address some of the most challenging problems faced by our naval forces. 



technical paper

Government agencies are rapidly accelerating AI across mission-critical operations to protect citizen wellbeing and national security. This momentum brings significant risks in predictive, generative, and agentic systems, including bias, opacity, safety failures, and governance gaps that must be managed.

This session highlights how government agencies can adopt a unified AI Observability and Security platform as their command center for AI deployments, giving them the oversight and audit trails they need to deliver high-performance AI, safeguard public trust, and maximize mission effectiveness and accountability. It enables teams to operationalize scalable, performant, and safe applications and models through continuous monitoring, diagnostics, and interpretability. 

We will share a case study on Fiddler’s work with the Defense Innovation Unit (DIU) on Project AMMO (Accelerated MLOps for Maritime Operations). Fiddler AI contributed observability capabilities for post-deployment monitoring, image embedding-drift detection with interactive root-cause analysis/explainability that accelerated Automatic Target Recognition (ATR) model retraining for unmanned undersea vehicles (UUV). Following the successful prototype, Project AMMO transitioned to production with the Naval Information Warfare Center (NIWC) Pacific.


AI Observability and Security: Your Command Center for National Security

TorchSig is an open-source, radio frequency (RF) machine learning (ML) library with synthetic signal datasets, domain transforms, and pre-trained models. This briefing will cover the v2.1.0 release, new pre-trained models, and new open-source repositories under TorchSig.
TorchSig's datasets are synthetically generated, which allows for the robust creation of large RF datasets for ML. In the v2.1.0 release, we have added even more modularity to our dataset and signal generators. Users now have the most flexibility and control over the generated signals. Users can 1) parameterize the datasets, signal family, or even specific signal level, and 2) create their own signals. For instance, a signal can be parameterized to only generate in a certain bandwidth of frequencies. Or, a completely new signal (e.g., Morse code) can be developed and easily integrated into the TorchSig signal generation pipeline. In addition to our 50+ already supported signal classes, users can adapt the library to their specific applications—whether digital, analog, or completely different signals. Under the hood, this means a restructuring of dataset and signal metadata, as well as slight API changes.
TorchSig now also supports users importing their own collected data into the ecosystem, which we call Bring Your Own Data (BYOD). This requires the user to write their own file handler so the library knows how to interface with their data. This feature also allows the user to augment their data with TorchSig's transforms and impairments.
We have released two pre-trained models under the torchsig-models repository. The first model is for narrowband classification, trained on IQ data with a dataset size of 1.14 million. The second model performs wideband energy detection on spectrograms, with a dataset size of 385k. These models are a good starting point for users, and can be fine-tuned on application-specific data. Note that both models are trained on TorchSig v2.0.0.
Lastly, we have released two more open-source repositories that use TorchSig. One is called gr-spectrumdetect, which uses a pre-trained v0.6.1 TorchSig wideband model with GNU Radio for live RF energy detection. Additionally, we have agentic-spectrumdetect, which combines a high-performance TensorRT-accelerated YOLO model and an LLM agent for agentic spectrum analysis.
TorchSig is developed by researchers at the Laboratory for Telecommunication Sciences (LTS) in College Park, Maryland. The team actively pushes regular releases on GitHub. We are always looking for feedback on how to make our library better for as many needs and applications. For this purpose, either a talk or workshop session would be beneficial to discuss how TorchSig can support the Navy's missions.

More information about TorchSig:
TorchSig website: torchsig.com
TorchSig GitHub page: https://github.com/torchdsp/torchsig

TorchSig: Open-source RF Machine Learning, v2.1.0 Release

To the selection committee: We would like to propose delivering this session in two formats--a 20 minute oral presentation and a hands-on workshop. The workshop format could be completed in 1.5-2 hours, so we would be happy to tailor it to a time frame that works for the conference. Please let me know if you have any questions.


Agenda:

How do you know if your data is behaving normally? Aerospace organizations face an ever-growing volume of time series sensor data. Detecting unusual patterns can be the key to improving performance or preventing failures. This session will explore techniques for time series anomaly detection, including statistical and AI-based approaches. Whether your focus is design, testing, or operations, you’ll learn how to uncover insights in sensor data that lead to more reliable and efficient systems.

Highlights include:

-Identifying anomalies in historical time series data
-Extracting and engineering features from time series signals
-Designing, comparing, and evaluating AI-based anomaly detection algorithms
-Deploying anomaly detection methods in real-time or streaming environments

Anomaly Detection Techniques for Time Series and Sensor Data

This tutorial/demo provides an introduction to the fundamentals of quantum computing from the software perspective, starting by demystifying the core concepts that differentiate quantum machines from classical ones, including qubits, superposition, and entanglement. Attendees will learn about quantum gates and how they form quantum circuits, highlighting the inherent differences in quantum algorithm design and the probabilistic nature of measurement. We will conclude with an application of quantum computing toward Machine Learning (QML), classifying simple data using a Variational Quantum Circuit (VQC). This provides a tangible glimpse into how these quantum principles are currently being researched for future applications to accelerate AI tasks.

Quantum Computing Fundamentals and Intro to Quantum Machine Learning

The "Co-Design for Edge AI" project is focused on developing and demonstrating an end-to-end co-design flow for integrating novel machine learning (ML) accelerators into specialized edge hardware optimized for mission size, weight, and power (SWaP) constrained applications. The project establishes tools and methodologies for identifying and implementing accelerators by analyzing application code to extract and convert computationally expensive portions into hardware description languages like Verilog. These accelerators target ML workloads that are currently performance bottlenecks, offering significant energy savings and enhanced mission capabilities compared to existing solutions. The co-design flow guides domain experts in concurrently designing hardware and software components, reducing design times from years to months, and enabling the rapid deployment of energy-efficient, specialized hardware for mission-critical applications such as UAVs, radio communications, and space satellite computing.
The project collaborates with various stakeholders to apply its co-design flow methodology across multiple application domains and partners. This approach is being transitioned to government and industry stakeholders, enabling early and informed hardware decisions. The project delivers application-specific accelerator designs that are significantly more energy-efficient than existing solutions, improving mission metrics by 20% and engaging multiple Department of Defense branches in the adoption and implementation of these advanced technologies.

Co-design for Edge AI

From biomedical sensing and monitoring to DVS camera output classification and RF fingerprinting, many applications require the ability to learn and adapt to spatiotemporal patterns, that is, patterns in multichannel time sequences. Spiking neural networks (SNNs) are bio-plausible computing models with high energy efficiency. The temporal dynamics of neurons and synapses enable them to detect temporal patterns and generate sequences. While Backpropagation Through Time (BPTT) is traditionally used to train SNNs, it is not suitable for online learning of embedded applications due to its high computation and memory cost as well as extended latency. In this work, we present Spatiotemporal Online Learning for Synaptic Adaptation (SOLSA), which is specifically designed for online learning of SNNs composed of Leaky Integrate and Fire (LIF) neurons with exponentially decayed synapses and soft reset. The algorithm not only learns the synaptic weight but also adapts the temporal filters associated to the synapses. Compared to the BPTT algorithm, SOLSA has much lower memory requirement and achieves a more balanced temporal workload distribution. Moreover, SOLSA incorporates enhancement techniques such as scheduled weight update, early stop training and adaptive synapse filter, which speed up the convergence and enhance the learning performance. When compared to other non-BPTT based SNN learning, SOLSA demonstrates an average learning accuracy improvement of 14.2%. Furthermore, compared to BPTT, SOLSA achieves a 5% higher average learning accuracy with a 72% reduction in memory cost. The extremely low computational cost of the learning algorithm allows us to perform both inference and training on the same FPGA platform, enabling true autonomy for mobile systems.

Neuromorphic Spatiotemporal Online Learning for Synaptic Adaptation

Illegal, unreported, and unregulated (IUU) fishing undermines maritime security and threatens the economic/political stability of coastal nations. For the U.S. Navy (hereafter, Navy) and allied forces, such activities represent not only an environmental or economic concern, but also a persistent intelligence and domain-awareness challenge that blurs the line between civilian and adversarial maritime behavior. Detecting, interpreting, and prioritizing these activities at scale requires automated systems that are both operationally reliable and analytically transparent.

This work presents an explainable, multi-stage machine learning methodology for large-scale detection and characterization of fishing activity using Automatic Identification System (AIS) telemetry provided by Global Fishing Watch (GFW). The system processes millions of vessel pings collected across varied ocean regions and time blocks to classify vessels as fishing or non-fishing, and—when applicable—to infer the likely fishing type (e.g., trawler, drifting longlines, purse seines, etc.). The model implements an XGBoost-based model template that processes various time-based statistics of AIS pings (e.g., instantaneous, daily, weekly, and historical AIS aggregations).

By using automated methods to identify fishing vessels based on their characterization, Navy operators can focus their attention on identifying vessels that are likely to be non-fishing. The examination of explainable results is key to making the best use of this system. This method integrates explainable AI (XAI) methods, including SHAP-based feature attribution, to help operators interpret model outputs and understand why vessels are identified/classified as they are. However, in isolation, these XAI methods have the potential to reduce operator efficiency, adding yet another set of data to interpret from the model. To mitigate this, we investigated the use of large language models (LLMs) to distill the model classifications, important features, and other XAI characteristics into a high-level, glanceable summary. These summaries enable operators to quickly assess the credibility of automated detections by highlighting key behavioral indicators such as motion irregularities, timing patterns, and spatial anomalies.

A complementary visualization interface, built on Folium, exposes these data and summaries, allowing analysts to rapidly explore regional patterns, identify anomalous behavior, and generate evidence-based assessments suitable for operational use. Ongoing evaluation is exploring the efficacy of the methods. This research contributes to the development of explainable, domain-aware machine learning pipelines for maritime situational awareness. Operationally, it provides investigations into a scalable, interpretable foundation for integrating AI-driven anomaly detection into Navy decision-support systems, advancing the broader goal of achieving efficient, trusted autonomy in maritime intelligence operations.

From Telemetry to Intelligence: An Explainable Pipeline for Fishing Activity Characterization

The ability to predict catastrophic system events is important for computers and autonomous systems overall, in order to employ mitigations or at least minimize damage. Predicting these “tipping points” using machine learning is one way to forecast when a system will destabilize and the time before expected failure. Frequently, the response time between detection and implementing a mitigation or shutdown needs to be at machine speeds vs. human-response speeds.
Remote missions and environments would by necessity need to leverage greater autonomy in their systems and respective preventive measures. Adversaries may initiate a non-conventional cyber-attack, assumed by the attacker to be undetectable until it is too late. Similarly, complex production / processing systems may be on the verge of failure for multiple reasons without the obvious clues available in time for a human operator to take action. This goes beyond normal cyber mitigations and respective system predictive maintenance, as there is a clear need to forecast such failures.
A JHUAPL – funded research effort called “Tip-it” has started to develop and test specific classes of algorithms to forecast critical events. These algorithms adapt to the system by internally constructing nonlinear mathematical models of the system and respective environmental dependencies. This allows them to predict behaviors several time steps before they occur. The first year of this research is concluding with the following results:
a) Successfully forecasted impending destabilizations for both small and large systems.
b) Worked on multiple scales ranging from a simple physical, to power grid-sized systems.
c) Leveraged theoretical (nonlinear mathematical equations), real physical systems, actual static production plant data, and simulated power grid data. 
d) Learned / forecasted simple physical systems chaotic orbital with zero-shot learning. This lightweight ML algorithm class successfully learned the dynamic chaotic orbital at first exposure- predicting both onset and length of chaotic orbitals. 
These results and associated technical details were presented in May 2024 to NASA LaRC, who has since expressed interest in the progress of this research.
The follow-on focus would be on a re-stabilization hypothesis based upon earlier research, which indicated that a few preselected perturbations can be added to a theoretical model in order to extend its stability beyond its normal performance envelope.
The end-objective would be twofold: a dynamically re-stabilizing system that is more resilient to attacks or failures, and a means of discretely de-stabilizing an adversary’s system.


Predicting Destabilizations in Complex Systems and Mitigations that Extend Performance

Underwater acoustic vector sensor (AVS) packages measure both acoustic pressure and particle velocity, enabling them to estimate bearings and elevations of acoustic signals without significant spatial aperture. AVS sensors play a crucial role in sonobuoy, towed array, and autonomous applications. While beamforming techniques are commonly applied to these sensor outputs, a range of additional nonlinear metrics can be derived from AVS using underlying physical principles. Such metrics include "azigrams," "elegrams," the kinetic-to-potential-energy ratio (K/E), polarization properties, and the normalized transport velocity (NTV), a measure that embeds information about the angular spread of an acoustic source. These metrics provide alternative time-frequency representations of data beyond a simple spectrogram and can be arranged as channels fed into an autoencoder (AE) for unsupervised learning and feature space clustering, similar to how camera images feed separate red, green, and blue channels into convolutional layers of an AE. We explore and demonstrate these concepts using hundreds of thousands of bowhead whale sounds collected in arctic waters between 2008 and 2014 and examine whether K/E and NTV metrics improve AE discrimination of angularly compact acoustic signals from both non-acoustic transients (e.g., knocks and vibrations) and transient signals from distributed sources (e.g., wind and reverberation).

Applying physics-based insights to unsupervised learning on acoustic vector sensors

Achieving optimal performance from machine learning (ML) algorithms in complex naval applications requires effective workload management.
This effort presents a novel ML Control network designed to enhance signal processing in challenging environments. The network intelligently routes data to specialized functions tailored for specific data characteristics, enabling optimized baseband extraction. We are currently investigating integrating this architecture into a Low Band Universal Communications System Receiver for submarine platforms, addressing the unique operational requirements of very low frequency (VLF) communications. The objective is to improve both bit error rate and anti-jamming performance while minimizing computational burden – a critical need for resource-constrained naval systems. This work seeks to streamline processing architectures and improve size, weight, power, and cost (SWAP-C) metrics, ultimately contributing to enhanced receiver designs for Naval Command, Control, and Communications (NC3) capabilities. We believe this research offers a promising approach to improving ML performance in demanding naval scenarios and invites collaboration with other researchers exploring similar challenges within the defense research community.


Premium content

Downloads

Next from NAML 2026

AI Observability and Security: Your Command Center for National Security

Similar lecture

AI Data integration & Automation For Maintenance

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES