Part of the DoW SBIR/STTR FY27 Release 2 topic list. Browse all 83 topics by industry or service.
Topic DON27TZ02-NV002 from the DoW STTR 2027 BAA, Release 2. Pre-release now; proposals due January 6, 2027, 12:00 PM ET.
Quick answer
DON27TZ02-NV002 is a Department of the Navy STTR topic seeking a new machine learning method that matches deep neural network accuracy while being explainable, able to recognize when it cannot give a reliable output, and efficient enough to run on low-cost CPUs and edge hardware. It is a Phase I STTR topic that requires a strong partnership with a research institution. Proposals are due January 6, 2027, 12:00 PM ET.
Solicitation at a glance
- Topic number: DON27TZ02-NV002
- Official title: Field-Friendly Super Intelligence and Machine Learning Systems
- Agency: Department of the Navy
- Program: STTR (DoW STTR 2027 BAA, Release 2)
- Award type: Phase I (see topic text for Phase II)
- Award amount: Phase I base up to $200,000 plus an option up to $115,000. Source: Navy Release 2 instructions in DSIP.
- CMMC level: Level 2 (Self)
- ITAR: No
- Technology areas: None
- Pre-release (direct TPOC contact allowed): November 4 to December 2, 2026, 8:00 AM ET
- Opens in DSIP: December 2, 2026, 8:00 AM ET
- Proposals due: January 6, 2027, 12:00 PM ET
- Official source: DSIP Topics app (search DON27TZ02-NV002)
BW&CO's take
What the Department of the Navy wants
The Navy wants, at minimum, an alternative to existing CNN-based architectures for image classification. The model must map parameter count to expected performance, make its input-output logic as readable as standard code, and recognize when it cannot give a reliable answer. Benchmarks are set against ResNet-18 and larger on MNIST, CIFAR-10, and CIFAR-100: equal or better accuracy with 50% or less of the parameters, training compute, training memory, and inference compute, plus sublinear growth. Phase I shows feasibility on image classification. Phase II runs the full experiment set and explains how the approach extends to transformers and multimodal LLMs and VLMs.
Who is a good fit
- A new model formulation that rethinks the basic problem, with early results on MNIST or CIFAR benchmarks
- Research on intrinsic interpretability that goes beyond post-hoc explanation techniques
- Methods for per-instance confidence or self-assessed reliability
- Experience running efficient inference on CPUs and embedded edge devices, and in mixed CPU-GPU settings
Watch-outs
- This is an STTR topic, and the text explicitly requires a strong partnership between the small business and a research institution.
- The topic notes that most known efforts modify the hardware and software of industry-leading NVidia systems and that none seems to rethink the basic formulation of the problem, which is the direction it describes.
- Not every metric is mandatory, but proposals are evaluated on the number of targets achieved against ResNet-18 and larger base architectures, so report results on all three datasets where possible.
- The topic requires CMMC Level 2 (Self). It is not marked as ITAR-restricted in the topic metadata.
Think your company fits DON27TZ02-NV002? Talk with BW&CO before the January 6 deadline.
Plan Your DON27TZ02-NV002 Submission
Official topic text (verbatim from DSIP)
The text in this section is copied word for word from the official DSIP topic as published on November 4, 2026. DSIP is the official source and topics can be amended. Always confirm the current version at dodsbirsttr.mil before you submit. Our commentary above is BW&CO's interpretation, not Government guidance.
Title
Field-Friendly Super Intelligence and Machine Learning Systems
Objective
Develop an Super Intelligence/Machine Learning (SI/ML) method that achieves the high-performance accuracy of modern Deep Neural Networks (DNNs) while mitigating their limitations.
Description
Current DNN architectures suffer from inherent "Black Box" vulnerabilities, demanding massive computational scaling and offering no clear link between parameter count and performance. The proposed model must address these shortfalls:
- Optimized Resource Utilization: Unlike DNNs that heuristically overparameterize to boost accuracy, the proposed model must quantitatively map parameter count to expected performance and enable users to identify the minimum configuration required for mission success, significantly reducing computational overhead, complexity, and deployment costs.
- Transparency: The model must replace opaque, distributed neural layers with novel methods for interpretability and trust that go beyond the current post-hoc techniques. By making input-output logic similar to standard programming code, the model should ensure decisions are easily explainable, verifiable, and compliant with auditing standards.
- Reliability: The model must recognize whether it can provide a reliable output for a given problem instance, which will address another well-known criticism of DNNs wherein performance is typically measured as an average, e.g., average accuracy, over many instances of a problem and will help build user trust and acceptability of the model vs DNNs, particularly in sensitive areas like person identification in a crowd video and identifying the source of a cyber-security violation.
The model must deliver DNN-level performance with superior speed and a reduced memory footprint, optimized for lean, secure processing on low-cost CPU and embedded edge hardware. By eliminating cloud-dependence, the model will maximize data privacy and operational security while substantially reducing power consumption compared to GPU-reliant models. It must operate flexibly across mixed CPU-GPU environments, allowing operators to dynamically balance throughput against computational cost. Beyond standard convolutional neural networks (CNNs), the model’s architecture must be extensible to transformer-based and multimodal models handling text, image, and video.
The metrics below establish outer-edge targets specifically for CNN-based image classification. While satisfying every metric is not mandatory, the model will be evaluated on the number of targets achieved and the viability of the engineering trade-offs.
- Comparison with Base Architecture of ResNet-18 and larger
- Comparison of the Number of Classes 10 and 100 using 3 datasets: MNIST, CIFAR-10 and CIFAR-100, having different combinations of numbers of classes and total number of image pixels (number of channels*number of pixels per image)
- Mean Average Accuracy of Classification: Same as DNNs or higher
- Number of Parameters (Fraction of Base Architecture): 50% of DNNs or less
- Training MFLOPs/Image (Fraction of Base Architecture): 50% of DNNs or less 6. Training Memory Size (Fraction of Base Architecture): 50% of DNNs or less7. Inference MFLOPs/Image: 50% of DNNs or less8. Rate of Growth of Metrics (5), (6), and (7): Sublinear
In Phase II, similar metrics and their extensibility and relevance to transformers and multimodality scenarios (images and videos as well as their combinations with text to form LLMs and VLMs) will need to be provided and their implementation explained.
This STTR topic requires combinations of research and development, and proposers must demonstrate a strong partnership between the small business and a research institution.
The proposed solution should eliminate the dependency on complex computing infrastructure (e.g., GPUs) by delivering comparable performance on standard off-the-shelf hardware (CPUs, edge devices). It should optimize the cost-to-performance ratio, balancing accuracy with reductions in time, storage, and computational cost. Furthermore, the architecture should ensure robust explainability—like conventional software programs—and provide reliable confidence metrics to establish user trust. At a minimum, the method should result in the development of an alternative to existing CNN-based architectures for image classification.
The many problems with the modern SI/ML architectures and models mentioned earlier are well recognized and reported. Many researchers and companies are currently attempting to address these shortcomings, including a number of startups. Most of the better-known ones are doing so by modifying hardware and software of the industry-leading NVidia systems. None seems to be rethinking the basic formulation of the problem which underlies the direction taken in the approach of the Model. If successful, the Model may turn out to have a major impact on the Naval systems and SI/ML industry which is already very large and growing fast. The deeper impact that could result from the lean design of Model could significantly add to the societal reach of SI/ML methods.
Phase I
Develop a minimal complexity design to assess the value of the concept of the model. (Note: The design does not have to be optimal but should serve as a feasibility to assess the art-of-the-possible, by conducting an adequate, convincing subset of experiments of those specified in the Description.) Conduct a feasibility study that should investigate all known options that meet or exceed the minimum performance parameters suggested in the Description and should also address the trade-offs and risks, in accordance with the level of innovation. Prioritize demonstrating feasibility on the core image classification challenge even though the long-term vision for the model is extensibility. Prepare a final Phase I report on the design and experimental results and a Phase II plan.
Phase II
Develop a minimal complexity model prototype and test it with at least all the experiments listed in the Description, ensuring that the results should clearly assess the absolute and relative (to DNN) performances of the scaled operational optimal prototype, and its scalability with increasing data complexity via the specified metrics, to demonstrate the model’s capabilities from the Phase I study determined to be most feasible. Provide the optimal model prototype; a report containing designs and testing results and how they demonstrate the power and future potential of the prototype; and a Phase III plan for prototype evaluation and other potential efforts as necessary. Provide metrics and their extensibility and relevance to transformers and multimodality scenarios (images and videos as well as their combinations with text to form LLMs and VLMs) and their implementation explained. Prepare a Phase III commercialization/transition plan.
Phase III Dual Use Applications
Generalize the CNN-targeted prototype model so it also applies to Transformers. Convert the prototype to an industry-friendly replicable form, implementable in one or multiple more progressive versions, as advised by industrial experts. Evaluate it on industry-grade large datasets of large images from very large numbers of classes. Deliver the final model prototype and a report containing designs and testing data. (Note: Detailed mission descriptions and effectiveness requirements will be addressed at a higher level of classification.)
Light-weight visual classification systems have many applications, particularly for autonomous robots.
References
Hadjis, S.; Zhang, C.; Mitliagkas, I.; Iter, D. and Ré, C. “Omnivore: An Optimizer for Multi-device Deep Learning on CPUs and GPUs.” arXiv:1606.04487, 2016. https://arxiv.org/abs/1606.04487
Han, S.; Mao, H. and Dally, W.J. “Deep Compression: Compressing Deep Neural Networks with Pruning, Trained Quantization and Huffman Coding.” 4th International Conference on Learning Representations (ICLR), 2016. https://arxiv.org/abs/1510.00149
Chen, B.; Medini, T.; Farwell, J.; Gobriel, S.; Tai, C. and Shrivastava, A. “SLIDE: In Defense of Smart Algorithms over Hardware Acceleration for Large-Scale Deep Learning Systems.” arXiv:1903.03129, 2019. https://arxiv.org/abs/1903.03129
Zhou, X.; Qu, X. and Pan, F. “A lightweight neural network model for security inspection targets.” 2022 IEEE 5th Advanced Information Management, Communicates, Electronic and Automation Control Conference (IMCEC), Chongqing, China, 2022, pp. 270-275. doi: 10.1109/IMCEC55388.2022.10019899
Pinto, D.G. and Mane, S.B. “Energy Efficient Implementation of CNNs Using Pruning: Techniques and Insights.” 2025 International Conference on Information, Implementation, and Innovation in Technology (I2ITCON), Pune, India, 2025, pp. 1-5. doi: 10.1109/I2ITCON65200.2025.11210640 https://ieeexplore.ieee.org/document/11210640
Keywords
Deep Learning Neural Networks; Black Box; Power consumption; Explainability; Trust; Edge Computability
Frequently asked questions
What is DON27TZ02-NV002?
It is a Department of the Navy STTR topic to develop a machine learning method that achieves the accuracy of modern deep neural networks while addressing their black-box behavior, heavy compute needs, and lack of per-instance reliability.
When is DON27TZ02-NV002 due?
Proposals are due January 6, 2027, 12:00 PM ET. DSIP pre-release began November 4, 2026, and direct contact with the topic authors is allowed only until the topic opens on December 2, 2026, 8:00 AM ET. After that, questions go through DSIP Topic Q&A.
Is this a Phase I or Direct-to-Phase II topic?
It is a Phase I STTR topic. Phase I is a minimal-complexity design and feasibility study focused on image classification, and Phase II builds and fully tests the prototype. A research-institution partner is required.
What benchmarks will the model be measured against?
The targets compare against ResNet-18 and larger on MNIST, CIFAR-10, and CIFAR-100. The model should match or beat DNN accuracy while using 50% or less of the parameters, training MFLOPs per image, training memory, and inference MFLOPs per image, with sublinear growth in those metrics.