Designing Efficient Domain-Specific Architectures for Autonomous Systems

Abstract

The rapid development of deep learning models is driving a remarkable expansion in capabilities for a wide array of real-world applications, from smart sensors to autonomous systems like self-driving cars and aerial robots. These innovations bring the promise of unparalleled intelligence and autonomy. Yet, efficiently implementing these AI models in autonomous systems poses a significant challenge, a key to unlocking their full potential in practical applications. As Moore's Law begins to plateau, computer architects are increasingly focusing on domain-specific architectures to meet the evolving performance demands of these complex domains. Designing domain-specific architectures for autonomous systems presents unique challenges. These systems are complex, involving multiple critical components such as compute systems, sensors, controllers, and physical limitations like size, weight, and power. This complexity is exacerbated by two main factors. First, current methodologies in designing domain-specific architectures often result in inefficiencies, as they focus narrowly on compute-centric metrics, neglecting the autonomous system's holistic performance needs. Second, the evolving landscape of AI models and the diversity of autonomous systems call for domain-specific architectures that are not only efficient in a holistic sense but also flexible enough to adapt to rapidly changing AI model landscape. This scenario underscores the necessity to develop methodologies and tools that span from efficient training of AI models to their characterization and the creation of automated design methodologies for domain-specific architectures, for the effective deployment of these models in autonomous systems. This thesis presents systematic methodologies and tools for designing domain-specific architectures. It introduces a holistic framework specifically crafted for training AI models for autonomous systems. It leverages deep reinforcement learning combined with domain randomization and hardware-in-the-loop techniques to validate AI models across various deployment scenarios. These methods ensure that the models are not only functional in simulations but also effective in revealing system-level bottlenecks when deployed on aerial robots. Furthermore, the thesis introduces tailored performance bottleneck tools like roofline models, designed for autonomous aerial robots. These tools are instrumental in identifying and addressing computational bottlenecks, while also considering sensor and physical characteristics unique to autonomous systems, thus optimizing system performance. Moreover, much of the research focuses on creating custom domain-specific architectures, employing machine learning as a tool to automate their design for autonomous systems. The thesis demonstrates that a cross-stack approach in designing hardware and software is critical for optimizing the safety and performance of autonomous systems. By integrating components such as sensors, compute elements, and controllers, this comprehensive strategy ensures that the domain-specific architectures are efficient and balanced, maximizing mission-level performance. Additionally, the thesis acknowledges the vast design space involved in creating domain-specific architectures and proposes standardized interfaces to apply machine learning automatic design space exploration. This approach streamlines the process, efficiently navigating and pinpointing optimal solutions, thereby significantly reducing the complexity and time required in the design process. In conclusion, the thesis contributes by providing a comprehensive methodologies, performance models and tools for the design and optimization of domain-specific architectures. These contributions not only address the current challenges in the field but also pave the way for future advancements in the deployment and efficiency of AI models in autonomous systems, ensuring their practical and effective application in a rapidly evolving technological landscape.