U.S. military researchers are looking to develop an automated software framework that can train machine learning (ML) models and organize them into levels of hierarchy. This new class of computer-generated design model has for aim to help engineers create and train artificial intelligence (AI) systems more rapidly and efficiently.
Modern ML algorithms are great stand-ins for real-world functions, however, they lack of meta-cognition and of composability. These algorithms store new information by training on additional data, but they don’t have any knowledge of real-world functions or systems and what they represent. Moreover, they are often deployed in isolation, making it even more difficult to use in real-life situations.
The Ditto: Intelligent Auto-Generation and Composition of Surrogate Models project searches to develop an automated software framework that will take in microelectronics system design, train ML design model of subsystem components, and allow then to expand and organized themselves into levels of hierarchy. By doing this, engineers will have the possibility to make more informed decisions early on in the design process, find defects quickly, and mitigate risks for critical applications.
The DARPA Ditto program thus tries to create an AI framework that can learn to generate and combine surrogate models for various components of a complex system all while maintaining and communicating surrogate accuracy and coverage, to finally integrate these models into one design.
This program will focus on simulating integrated circuits (ICs), mixed-signal circuits boards, and networked-distributed systems in order to optimize the framework so it can adapt continuously to more designs and learn from mistakes. The frameworks will address one of these three different system design types.
The Ditto project will first develop a framework that demonstrates the functional capabilities by applying third-wave AI techniques to create surrogate models automatically. Then, it will develop a proof-of-concept framework with meaningful performance gains in a full-system simulation.
