| KEYWORD |
Innovative Neural Architecture Search for Reliable Neural Networks on MCUs
keywords ARTIFICIAL INTELLIGENCE, CONVOLUTIONAL NEURAL NETWORKS, DEEP NEURAL NETWORKS, FAULT INJECTION, FAULT TOLERANCE, TESTING
Reference persons DANIELE JAHIER PAGLIARI, ANNACHIARA RUOSPO, EDGAR ERNESTO SANCHEZ SANCHEZ
External reference persons ALESSIO BURRELLO
Research Groups DAUIN - GR-05 - ELECTRONIC CAD and RELIABILITY GROUP - CAD, DAUIN - GR-06 - ELECTRONIC DESIGN AUTOMATION - EDA
Thesis type EXPERIMENTAL, SOFTWARE DEVELOPMENT
Description Deep neural networks have demonstrated enormous potential in processing complex data and solving problems in various industries. However, the reliability of such networks in safety-critical applications has become a critical issue. This thesis proposal focuses on the exploration of innovative methods of neural architectural search of deep neural networks, which take into account both accuracy and reliability, considering the controlled injection of faults during training.
Goals:
The main objective of this thesis is to develop innovative approaches to simultaneously improve the accuracy and reliability of deep neural networks while reducing their latency.
Specific goals include:
- To evaluate the effectiveness of deep neural network architectures using different fault injection techniques in terms of model accuracy and reliability;
- To compare various architectures of deep neural networks obtained thanks to neural architectural search algorithms that do not consider the reliability in its optimization function in terms of accuracy and reliability of the models;
- To propose methods and strategies for the design of deep neural networks that optimally balance accuracy and reliability, considering the presence of faults in the real world, and reducing their size.
Proposed methodology:
The research will be conducted through an experimental approach which includes the following phases:
- In-depth literature search to gain a comprehensive overview of current deep neural network architectural research techniques and issues related to accuracy and reliability;
- Study and identification of possible sources of failures or malfunctions that may occur during the deployment of deep neural networks;
- Study of controlled fault injection techniques, which allow introducing specific faults during the training of neural networks;
- Implementation and evaluation of deep neural network architectures coming from classical NAS algorithms using different benchmark datasets and fault injection techniques;
- Proposal of methods and strategies for deep neural network architecture search that take into account both accuracy and reliability.
Required skills Required skills include C and Python programming. Further, a basic knowledge of computer architectures and embedded systems is necessary. Knowledge of deep learning and the corresponding models is also required. It is also useful (but not strictly necessary) to have a minimum familiarity with testing and fault tolerance concepts.
Deadline 30/12/2022
PROPONI LA TUA CANDIDATURA