Introduction:

In a previous blog post, we discussed some recent success Theta has achieved with our prototype AI model named ‘Naissance’ which is trained to detect the T-72 Main Battle Tank within remote sensing imagery:

Press Release – Naissance AI Model Prototype for detection of the T-72 Main Battle Tank

‘Naissance’, coming from the Latin verb nasci as in the word ‘nascent’, represents an idea which has recently come into existence. This model was an important first step for to test the viability of machine learning and AI for Aided Target Recognition (ATR) of high value objects within digital imagery. The experiment showed that a high degree of accuracy and specificity could be achieved given the right model architecture and a large enough dataset.

Through a revolution in Neural Network architecture, our firm has developed a next-generation model named ‘Renaissance’ which greatly exceeds the performance of those which came before it. The full process for its development following the older ‘Naissance’ model is described below.

---

Theta publishes software called OpenAthena™ which allows instant calculation of ground coordinates from any pixel of any single drone image. The user may simply tap anywhere within a drone image to obtain the corresponding ground location. The location can then instantly be shared over a network to devices running ATAK.

The software introduces a novel terrain-raycast technique for processing of remote sensing data from drones. This unique approach has significant advantages in speed and simplicity for processing drone imagery, removing the need for multiple images and intensive post-processing computation.

For the software operator, the benefits of this technique manifest in a system which is faster and more accurate to use for obtaining actionable location information from drone imagery. These benefits are crucial for time-critical applications such as search and rescue, law enforcement, and defense, where speed, simplicity, and accuracy can mean the difference between a successful operation and an unsuccessful one.

Role of Artificial Intelligence:

At present, OpenAthena still requires a human operator to manually-identify and select an object of interest within drone imagery. This is perfectly adequate for processing imagery from a small area, but can pose significant difficulty when a human needs to review hundreds or thousands of images covering much larger areas. Humans need to be trained to recognize different classes of objects, they get bored, tired, can miss things, and require breaks and shift turnover for continuous operations. Additionally, each human reviewer operating near the front lines of conflict requires their own significant logistical tail. AI object detection models like Naissance for ATR, by contrast, can process thousands of images very quickly and thoroughly, operating 24/7 while never getting tired, bored, or needing a break.

The adoption of object detection AI for ATR in military contexts is expected to drastically shorten each cycle of the Observe, Orient, Decide, and Act (OODA) loop identified by former Air Force Colonel John Boyd as critical to battlefield outcomes. Autonomous systems may perform the Observe and Orient tasks of this loop instantaneously, speeding up the portion of the cycle prior to humans’ own Decision and Action. The combatant who can complete each full cycle of this OODA loop faster than their adversary is expected to enjoy a ‘decision advantage’ whereby they may routinely disrupt and overwhelm the enemy before they have the opportunity to react.

Limitations of AI:

No matter how well they perform, such AI systems are not perfect. In rare cases, they can make mistakes (such as false positives or false negatives) of the kind that would never occur with a human analyst. In life or death situations, a human review is always needed due to the inherent unpredictability of such systems.

Figure 3 (above) illustrates such a unusual and unpredictable failure case which was encountered early in the Naissance model’s development. The model routinely predicted that chicken coops were T-72 tanks with high confidence. Due to the model’s reliance on internal hidden layers for coming to a determination, it’s impossible to know why it came to it.

With the addition of more training data this behavior disappeared entirely, however its emergence in the first place is illustrative of the technical and ethical challenges of the use of AI for ATR. Deep learning techniques fundamentally are unable to provide the predictability and accountability needed for life or death decisions, especially compared to a human. They are much better suited however to the Observe and Orient tasks, helping humans to find the ‘needle in the haystack’ of copious data which can then be Decided and Acted upon.

Initial testing for project Naissance:

The original training and testing set for the Naissance model was a large dataset of synthetic computer graphics imagery. With such a sufficiently large dataset, the model achieved nearly perfect performance for detecting the desired object class:

Figure 4
Test results from Naissance model on 932 labeled images never seen before by the model

Naissance robustness testing:

These tests were performed under ideal conditions. For the purpose of our research however, we desired to test how well this model would perform on complex real world scenes. The results however were much different than the above test:

Figure 5
Test results from Naissance model on a labeled, real world dataset
The Confusion Matrix, Recall curve, and Precision/Recall Curve exhibit
significantly degraded performance compared to the previous test

The sub-optimal results of the previous test prompted our team to transition to an experimental, next-generation state of the art object detection AI model architecture.

Renaissance AI

Compared with previous architecture generations, this architecture contains an auxiliary branch which significantly reduces the distortion of visual information provided to the objective function throughout the model.

With this new architecture, our observation has been that its performance is drastically improved in multiple categories:

• The model learns much more quickly. Loss function convergence occurs much sooner in the training process, achieving the same level of mean Average Precision (mAP) within fewer training epochs.
• Significantly reduced rate of false positives, especially in complex scenes
• Significantly improved rate of recall, even in difficult and noisy scenes

Given these significantly improved results, we named this new model ‘Renaissance’. This name reflects its status as a rebirth of the original concept for project Naissance.

The test results and examples of the models significant capabilities are provided below:

Figure 5
Test results from new ‘Renaissance’ model on a labeled, real world dataset
The new Recall curve and Precision/Recall Curve exhibit
significantly improved performance compared to previous architectures

Notably, the Renaissance AI model demonstrates significant ability to recognize the target object class even when its key identifying features are heavily obscured. This property is advantageous compared to simpler algorithmic approaches traditionally used for ATR:

Video Demo:

Methodology:

The Naissance and Renaissance projects are based on variants of the YOLO object detection framework. Proprietary techniques were developed by our firm for the curation, labeling, and transformation of data as well as the training process itself.

Just as in the Naissance project, the training data for project Renaissance was labeled and curated by hand over a significant quantity of hours of labor. We developed a procedure to issue secure computer hardware to U.S. based independent contractors for assistance in data labeling.

For model training, we transitioned away from the use of two NVIDIA RTX 4070 12Gb GPU’s on a Linux Workstation. Instead we designed and built a dedicated High Performance Computing (HPC) server named “ZEUS”, based on entirely consumer-grade PC Hardware:

ZEUS HPC parts list:

Component	Model Name
Motherboard	MSI PRO B550-VC
GPU(s)	4x Zotac (NVIDIA) GeForce RTX 3090 24GB Trinity OC
CPU	AMD Ryzen 5 5500 6 Core 3.6 GHz
RAM	128 Gb G-Skill Ripjaws V DDR4-3200
SSD	Samsung 980 Pro NVMe
PSU	Seasonic Prime TX-1600 | 1600W | 80+ Titanium
Case	Kingwin KC-6GPU-SS open air miner case
PCIe extenders	4x Asiahorse 200mm PCIe 3.0 x16 shielded extender

Performance:

142 Trillion Floating Point Operations Per Second (FLOPS) at 16-bit precision

Tuning:

Traditional performance tuning operations on the Linux operating system improved general stability and performance. To prevent the computer from drawing too much electricity and tripping an electrical breaker, we adapted a script from the following article which limits the power usage of the NVIDIA GPUs (with only modest performance impact):

Quad RTX3090 GPU Power Limiting with Systemd and Nvidia-smi

https://gist.github.com/mkrupczak3/2c7f17d7730b11c981b7feab48e76d4c

Compared to the use of cloud hardware rental or NVIDIA’s datacenter hardware solutions, the use of consumer hardware allows us to operate at significantly lower marginal cost for the production of each AI model. This approach significantly improves the rate at which we can iterate for production of new AI models.

Conclusion:

The pace of emerging technology in aerial robotics, data processing techniques, and AI is rapidly precipitating changes in land warfare doctrine. Our firm, Theta Informatics LLC, aims to be the nexus of progress of such technology in order to ensure free societies are better equipped and prepared for it than autocratic adversary nations. Autocracies’ inflexible thinking, poor tolerance of diverse groups and ideas, and rigid protocols will prevent them from adopting such technology as efficiently as free societies. Doctrinal evolution is necessary to preserve the rules-based international order established by free societies since the end of WWII in the face of historic challenges this century.

At the time of writing, Theta Informatics LLC is a pre-seed startup specializing in software and intellectual property holdings. We are actively looking for strategic partners including UAS manufacturers, system integrators, software contributors, investors, and end users in service of our firm’s mission.

We can be reached at the following email: inquiries@theta.limited

To see more from Theta, sign up for our newsletter and follow our official X (formerly Twitter), YouTube, and LinkedIn pages.