Strojové učení zohledňující hardware: Od automatizovaného návrhu k inovativním a vysvětlitelným řešením

Strojové učení zohledňující hardware: Od automatizovaného návrhu k inovativním a vysvětlitelným řešením

grant

evolutionary algorithm;approximate computing;deep neural network;machine learning;hardware accelerator;explainability;design automation;

As machine learning (ML) technology penetrates embedded devices, a new class of design automation algorithms capable of generating hardware-aware implementations of ML algorithms is highly desired. In addition, a lot of effort is now invested in developing explainable ML. We hypothesize that the design time of hardware-aware implementations of ML systems showing additional properties (such as explainable behavior) can be substantially reduced if the used design automation algorithms employ suitable surrogate models for estimating the accuracy, hardware parameters, and other desired properties. In addition to developing suitable surrogate models, we will create a new method based on genetic programming for the automated design of highly-optimized ML models showing excellent trade-offs among the quality of service, hardware parameters, and explainability. The design method and ML models automatically generated by the method will be evaluated in case studies, including image classifiers, Parkinson's disease assessment, and command classifiers of brain signals.

Sekanina Lukáš, prof. Ing., Ph.D. (UPSY FIT VUT) , research leader
Hurta Martin, Ing. (UPSY FIT VUT)
Malik Aamir Saeed, Ph.D. (UPSY FIT VUT)
Mrázek Vojtěch, Ing., Ph.D. (UPSY FIT VUT)
Piňos Michal, Ing. (UPSY FIT VUT)
Vašíček Zdeněk, doc. Ing., Ph.D. (UPSY FIT VUT)
Zaheer Muhammad Asad (UPSY FIT VUT)