Application deadline: 01/09/2023
Robust and Efficient Edge-AI for the Analysis of Images on Satellites (E=(AI)²)
Supervisor: Prof. Chiaberge, S.Angarano
Skills: Deep Learning, Python
Description: Research Grant (12 months) for a project funded by the Italian Space Agency (“ASI Research Day 2020 – Onboard AI” call).
Abstract: The project aims to develop AI methodologies for the onboard processing of optical data and images and to implement the designed Edge-AI system on a hardware platform with power consumption suitable for use in space. The project will see the collaboration between Politecnico di Torino and the companies Argotec and Ithaca. In particular, Politecnico is responsible for defining the requirements and designing a multitask and robust Edge-AI model for image analysis.
The work package (WP) on which the research fellow will work concerns the development of methodologies for verifying the integrity of neural network parameters used for image analysis in the Edge-AI system. The idea is to avoid designing inherently robust neural networks for two reasons: 1) the robustness obtainable architecturally with appropriate training is never perfect or guaranteed, and 2) the networks thus developed pay for the acquired robustness with degradation in image analysis performance (e.g., the accuracy of a classifier). This WP will analyze approaches based on using a neural network not specifically designed to be robust, coupled with methodologies for the timely detection of parameter errors and their correction.
Initially, a review of the state of the art related to error-robust neural networks will be conducted. Reference methodologies will also be defined based on the characteristics of the target hardware defined by previous WPs. The simplest approach is to compare the parameters of the networks in the memory used to do computations with those in nonvolatile memories; however, this approach requires heavy access to both the ROM and RAM of the system. Alternative blind methods, which rely on observing network inputs and outputs but not the parameters themselves, will therefore be studied, and a comparative analysis of the blind approach with the benchmark methods will be made, highlighting the advantages and disadvantages of each. The output of the task will be the definition of an integrity verification methodology suitable for satellite Edge-AI and compatible with the target hardware and its implementation in a high-level language.