Search academic texts

Information × Registration Number 0222U005297, 0121U109787 , R & D reports Title Development of new mobile information technologies for person identification and object classification in the surrounding environment popup.stage_title Head Koriashkina Larysa , Кандидат фізико-математичних наук Registration Date 27-12-2022 Organization National Technical University Dnipro Polytechnic popup.description2 For the first time, a model of a post-train adaptive block for a convolutional neural network is proposed. It is based on an adaptive selector with several branches of inverted residual blocks. Also, we present a method of training such a network based on PTA configuration sampling during training, which allows to train the network once, and then dynamically change the inference time of the network and its quality for efficient execution on devices with significantly different computing power. A new method for accelerating the adaptation phase of optimization meta-learning algorithms based on lambda patterns has been developed, which allows to significantly accelerate the adaptation phase with minimal loss of quality. Moreover, the  proposed method increases the quality of the one-step adaptation phase of meta-learning. For the first time, a model of a mobile access control system based on RFID tags has been proposed, which allows to significantly simplify the implementation of access control systems at enterprises, and to increase its security compared to the most common access control systems based on RFID access cards due to additional verification of the user's face and an anti-spoofing system. The MobileNetV2 neural network received further development thanks to the introduction of post-train adaptive blocks to adaptively change the network configuration after the training procedure is complete. The configuration change is performed taking into account the power of the target device. Also, the Model-Agnostic Meta-Learning optimization algorithm was improved, thanks to the introduction of Lambda-patterns to increase speed and quality of the meta-learning adaptation phase. Product Description popup.authors popup.nrat_date 2022-12-27 Close