Methodology for evaluating model-accelerator co-design in deep learning

Context

Deep Learning (DL) refers to a machine learning method of artificial intelligence in which neural networks approximate any functions and solve a wide variety of tasks. In most cases, they achieve higher prediction accuracy than humans. However, they are rarely as efficient as humans. In order to increase e.g. energy efficiency, special hardware accelerators are designed, which, however, often degrade the accuracy by applying compression methods. In this context, the evaluation of the accuracy resource trade-off is a major challenge.

Goals

This thesis aims to develop a methodology for evaluating the accuracy resource trade-off. First, the work starts with a literature review and identification of suitable concepts for the combined evaluation of the prediction accuracy of the DL model and the hardware resources and costs used, e.g. energy efficiency, latency, bandwidth. Based on this, a proprietary methodology is developed that can be used to evaluate the model-accelerator co-design. Its suitability will finally be tested in the evaluation of existing accelerators.

Requirements

• Programming experience in Python and C++
• Basic knowledge of neural networks
• Motivation and interest in solving technical problems independently