Hardware Implementation of Early Exits in Hardware Accelerators

Context

Early exits prove to be a capable method for adapting neural networks to changing execution contexts. While branching neural networks and early exits are actively researched methods, the implementation of these methodologies into hardware accelerators is rarely undertaken. Such implementation is vital for real-world use and needs to be done in order to integrate them into existing embedded systems.

Tasks

In this work, the student is assigned the task of implementing a methodology and generation capability to incorporate early exits into designated AI accelerators, such as Gemini.

Requirements

• Good knowledge in hardware development using HLS languages
• Good knowledge in machine learning methods and models