Embedded Systems

Analyzing the robustness of resource-efficient DNNs

Master’s thesis

Description

Deep neural networks (DNNs) are known to perform exceptionally well on a variety of different tasks but require substantial computational resources to operate. This is a major problem for using DNNs in real-world applications, e.g. an autonomous driving agent should require the least possible resources for navigation to reduce fuel costs. Thus, over the course of years, many techniques have been proposed to compress DNNs without performance loss, such as pruning, quantization, via knowledge distillation and many others.

Commonly, the performance loss for these techniques is gauged on the same distribution the DNNs have been trained on. However, DNNs struggle if the distribution changes between training and testing. For autonomous driving, it means that an agent will struggle to navigate in rainy or snowy conditions if they were not part of the training data. To the best of our knowledge, the influence of model compression on the robustness of the model has not been studied systematically so far.

In this work, we want to investigate whether state-of-the-art model compression methods retain model performance when tested on data that has not been part of the training data, i.e. whether compressed models are as robust as their uncompressed counterparts. We believe this question is crucial to pave the way for using DNNs in real-world applications: e.g. an autonomous agent needs to be both resource-efficient and robust to unforeseen conditions.

Many methods have been proposed to increase model robustness to distribution shifts; a different direction or an extension of this work might be an evaluation of the efficacy of state-of-the-art robustification methods for compressed DNNs.

ToDos

  • Research, implement and test different state-of-the-art compression techniques for DNNs.
  • Evaluate the robustness of these techniques on common robustness benchmarks for computer vision.
  • Alternatively or as an extension: Evaluate the efficacy of state-of-the-art robustification methods on compressed models.

Who

  • Master Students with interest in Deep Learning.

Prerequisites

Knowledge of Pytorch and Python is highly beneficial.

Where

In the Robust Perception team of the Embedded Systems chair, https://embedded.uni-tuebingen.de/team/

Contact

Bringmann, Oliver

Rusak, Evgenia