Machine Learning determines rules or behaviors from data that can be generalized and applied to new data. Applications of machine learning are growing rapidly—benefitting from increasingly affordable computation and a proverbial data deluge—but the raw data that a machine learner needs is often ill-suited for direct input. Raw data can be grammatically inconsistent, misaligned, or otherwise poorly situated in a way that renders learning results unusable. For technically savvy data scientists, many DIY solutions exist to address these problems. However, choosing, customizing, and creating a pipeline for your specific data type(s) and use case pose substantial challenges.
Our shape analytics approaches extract informative features and reduce the computational complexity and dimensionality of non-traditional or high-dimensional data types, expanding the utility of machine learning approaches. For applications with multiple input data sources, such as sensor networks or biological datasets, our data fusion approaches combine multiple data types and improve the results of machine learning.
GDA brings practical know-how and clever research to apply machine learning effectively and reproducibly. Our team draws on diverse research backgrounds to extend existing algorithms, offering novel unsupervised and reinforcement learning in addition to common supervised approaches. Practically, GDA can warehouse, catalog, and stream data to scalable learning pipelines.
Within general machine-learning, GDA has specialities in reinforcement learning, distributed machine-learning, and validation & verification.
Many applications require answers to how a system should act or react to information to achieve a goal. Reinforcement learning (RL) is a collection of techniques to develop optimal actions or controls in the face of limited information. Such collections of optimal actions are called a policy. Using deep neural networks to approximate policies has opened new applications up to reinforcement learning—more descriptively deep reinforcement learning. RL solutions are complex, computationally intensive, and often ill-defined.
GDA applies disciplined software development and research capabilities to extend existing reinforcement learning algorithms and practices to address a variety of multi-agent problems (e.g. controlling a group of UAV or sensors). Our team of experts create efficient computing solutions and work closely with subject matter experts to define and refine optimal solutions to RL problems in decision support models, power distribution, and cybersecurity.
Distributed machine learning
In many modern IoT applications, processing of input and subsequent inference happens in different locations, termed distributed machine learning. Typically, the former is “on the edge” and the latter is “in the cloud.” Specific mission conditions often dictate that one must be very mindful of how much information is transmitted between edge and cloud. These communication and computation constraints can limit the application of machine learning approaches.
GDA’s cutting-edge work on informative compressed representations and featurizations, based in shape analytics, allow information-rich summaries to be computed on the edge and powerful machine learning to take place in the cloud. This facilitates essential optimization of storage, processing, transmission, and learning.
Verification and validation systems
Every AI system performs well on its training data. Many schemes exist to verify that training success translates to success in reality, but these verification schemes are limited by imperfect foresight. Once deployed, an AI system is guaranteed to encounter unforeseen circumstances. Without an intelligent response to new and ambiguous inputs, even a verifiably well-trained AI system can make a catastrophic mistake.
GDA performs fundamental research in designing verification and validation systems (V&V) for AI, with the goal of providing certificates of novelty robustness for specific AI systems. Our V&V toolset identifies blind spots and corner cases in the data-, feature-, and decision-spaces of an AI system, giving users a proactive approach to catastrophe mitigation.