Delivering AI-Enabled Autonomy
At Scientific Systems, we develop technologies that enable autonomous systems to acquire information to fill knowledge gaps, predict possible outcomes, and choose the best course of action. As experts with real world experience, we have algorithm expertise in supervised and unsupervised learning to handle data-rich learning problems.
By integrating research from multiple fields including information theory, control theory, statistical learning, optimization, and game theory, we enable autonomous systems to better plan, act and react in uncertain and adversarial environments.
Transform Data to Knowledge
We are experienced in extracting knowledge from diverse data sources for optimal inferences and decision-making. We have developed statistical methods and artificial neural networks-based solutions for big and “small” data problems. The sources of data in our applications range from sensor data such as EO/IR, RF, GPR, acoustics, SAR/ISAR and GMTI, to human intelligence such as field reports and social media feeds.
Depending on the nature of the data and the constraints of the system, we develop problem-specific solutions from algorithm building blocks and their innovative extensions, such as: manifold learning, decision trees, Bayesian non-parametrics, and deep learning. Solutions delivered to our customers include landmine detection, dismount recognition and tracking, aircraft fault detection, gunshot detection and localization, and sniper recognition.
Integrate Domain Knowledge & User Insight
We integrate experts’ domain knowledge to build smarter generative models to fill data gaps or mitigate poor data quality. What’s more, we pair users with machines during on-line operations to reduce uncertainty. At Scientific Systems, we leverage this knowledge to create real impact. We assimilate expert insight in demographics and seasonal fluctuations to predict flu epidemics, prior oceanographic survey for seafloor mapping, human auditory processing for gunshot detection and localization, and more.
During operations, teaming users with machines reduces uncertainty. We incorporate human perception and tactical knowledge to improve automatic target recognition performance, and to disambiguate adversary tactics during real-time operations. Based on our expertise in state space modeling, Markov random fields and various statistical models, as well as our experience in computational schemes such as MCMC and Genetics Algorithms, we have developed approaches to incorporate domain knowledge and user insights that are compatible with customers’ ConOps and real-world constraints.
Optimize Information Acquisition
Performance of machine learning is closely tied to the quality of data provided to the algorithms. We design our AI-driven autonomous systems to actively acquire information that maximizes mission performance. Our closed-loop sensing technologies, taking environmental factors and machine learning algorithm performance into account, inform autonomous systems to acquire data that effectively improve the target knowledge.
This synergistic integration of machine learning and autonomy has resulted in significant performance improvement in object recognition, RF emitter localization, and sensor fusion.
Operate in Adversarial Environments
We have developed action plans to operate in uncertain and adversarial environments. Many of our applications are dynamic in nature, even with only partial and imperfect environmental information available. We’ve also developed AI-driven planning schemes to explicitly handle information deficiency as well as intelligent adversarial response.
At Scientific Systems, we employ concepts and tools from POMDP, Game Theory, and Reinforcement and Inverse Reinforcement Learnings to plan optimal future actions amid uncertainty. So far, we have successfully developed plans for tracking evasive underwater vehicles, fighter aircraft combat tactics learning, and strikes against IADS.