Publicado em: 06/12/2021
Prof. Reis efetua tutorial no IEEE ICECS 2021
O Prof. Ricardo Reis efetuou tutorial no IEEE ICECS 2021, 28th IEEE International Conference on Electronics Circuits and Systems, conferência flagship da IEEE CASS, que aconteceu de 28 de Novembro a 1º de Dezembro, em Hotel Sofitel Dubai The Obelisk, em Dubai, UAE.
O Tutorial teve uma duração de 3 horas, sendo que a primeira parte foi efetuada pelo Prof. Ricardo Reis e a segunda parte pelo Prof. Luciano Ost, da Loughborough University (Inglaterra). A primeira parte focou em IoT e Avaliação e Mitigação de Soft Errors a nível de circuitos e na segunda parte, o foco foi em Avaliação e Mitigação em alto nível de abstração do projeto,
Titulo: Deployment and Soft Error Assessment of Deep Inference Networks in Resource-constraint IoT Devices
Abstract: Deep neural networks (DNNs) are being incorporated in resource-constrained Internet of Things (IoT) devices, which typically rely on reduced memory footprint and low-performance processors. To enable DNN models’ execution on the underlying devices, software libraries and application programming interfaces (APIs) have been proposed. Such libraries/APIs are devoted to streamlining the design and development of embedded deep learning-based ap- plications through the fine-tuning of pre-trained network models, thus enabling their efficient execution in edge-computing platforms. For the time being, the majority of embedded trained models and their inference engines have been evaluated only according to their accuracy and performance over a given dataset. With the growing adoption of DNNs in safety-critical em- bedded systems (e.g., autonomous vehicles), increases the demand for safe and reliable mod- els. To reach levels of reliability that are comparable to those required by high safety stan- dards, it is imperative to supply electronic computing systems with appropriate mechanisms to reduce their vulnerability to radiation-induced soft errors. In this talk, we will discuss the key challenges, benefits, and impediments of deploying and assessing the soft error reliability of deep inference networks in resource-constraint IoT devices, considering different case studies covering different libraries/APIs and DNN characteristics.