Tuning small-sized convolutional neural networks for image classification using a genetic algorithm
Jazyk: | ruština |
---|---|
Rok vydání: | 2022 |
Předmět: |
convolutional neural networks for image classification
hyperparameter tuning генеÑиÑеÑкие алгоÑиÑÐ¼Ñ Ð½Ð°ÑÑÑойка гипеÑпаÑамеÑÑов обÑÑение нейÑонной ÑеÑи genetic algorithmÑ architecture selection Ð¿Ð¾Ð´Ð±Ð¾Ñ Ð°ÑÑ Ð¸ÑекÑÑÑÑ ÑвеÑÑоÑнÑе нейÑоннÑе ÑеÑи клаÑÑиÑикаÑии изобÑажений neural network training |
DOI: | 10.18720/spbpu/3/2022/vr/vr22-2526 |
Popis: | Тема вÑпÑÑкной квалиÑикаÑионной ÑабоÑÑ: «ÐаÑÑÑойка малоÑазмеÑнÑÑ ÑвеÑÑоÑнÑÑ Ð½ÐµÐ¹ÑоннÑÑ ÑеÑей клаÑÑиÑикаÑии изобÑажений Ñ Ð¸ÑполÑзованием генеÑиÑеÑкого алгоÑиÑма». ÐÐ°Ð½Ð½Ð°Ñ ÑабоÑа вÑполнена Ñ ÑелÑÑ Ð¸ÑÑÐ»ÐµÐ´Ð¾Ð²Ð°Ð½Ð¸Ñ Ð¿ÑÐ¸Ð¼ÐµÐ½ÐµÐ½Ð¸Ñ Ð³ÐµÐ½ÐµÑиÑеÑÐºÐ¸Ñ Ð°Ð»Ð³Ð¾ÑиÑмов в ÑвеÑÑоÑнÑÑ Ð½ÐµÐ¹ÑоннÑÑ ÑеÑÑÑ ÐºÐ»Ð°ÑÑиÑикаÑии изобÑажений, в ÑаÑÑноÑÑи, ÑаÑÑмоÑÑено пÑименение в ÑледÑÑÑÐ¸Ñ Ð¾Ð±Ð»Ð°ÑÑÑÑ :-     наÑÑÑойка гипеÑпаÑамеÑÑов нейÑоÑеÑи,-     обÑÑение полноÑÑÑÑ Ð½ÐµÐ¾Ð±ÑÑенной ÑеÑи,-     ÑиниÑное обÑÑение нейÑоÑеÑи,-Â Â Â Â Â Ð¿Ð¾Ð´Ð±Ð¾Ñ Ð°ÑÑ Ð¸ÑекÑÑÑ Ð¼Ð¾Ð´ÐµÐ»Ð¸. Ðомимо ÑÑого, пÑедложен ваÑÐ¸Ð°Ð½Ñ ÐºÐ¾Ð¼Ð±Ð¸Ð½Ð°Ñии наиболее Ñ Ð¾ÑоÑо ÑÐµÐ±Ñ Ð¿Ð¾ÐºÐ°Ð·Ð°Ð²ÑÐ¸Ñ Ð²Ð°ÑианÑов пÑÐ¸Ð¼ÐµÐ½ÐµÐ½Ð¸Ñ Ð³ÐµÐ½ÐµÑиÑеÑкого алгоÑиÑма. Ð Ñ Ð¾Ð´Ðµ вÑÐ¿Ð¾Ð»Ð½ÐµÐ½Ð¸Ñ Ð±Ñли ÑаÑÑмоÑÑÐµÐ½Ñ Ð¾ÑновнÑе пÑинÑÐ¸Ð¿Ñ Ð¿Ð¾ÑÑÑÐ¾ÐµÐ½Ð¸Ñ Ð¸ ÑабоÑÑ ÑвеÑÑоÑнÑÑ Ð½ÐµÐ¹ÑоннÑÑ ÑеÑей, пÑоизведен Ð¾Ð±Ð·Ð¾Ñ ÑÑÑеÑÑвÑÑÑÐ¸Ñ Ð°ÑÑ Ð¸ÑекÑÑÑ ÑвеÑÑоÑнÑÑ Ð½ÐµÐ¹ÑоÑеÑей пÑедназнаÑеннÑÑ Ð´Ð»Ñ ÐºÐ»Ð°ÑÑиÑикаÑии изобÑажений. Также вÑполнен Ð¾Ð±Ð·Ð¾Ñ Ð³ÐµÐ½ÐµÑиÑеÑÐºÐ¸Ñ Ð°Ð»Ð³Ð¾ÑиÑмов и ÑÑÑеÑÑвÑÑÑÐ¸Ñ Ð½Ð° даннÑй Ð¼Ð¾Ð¼ÐµÐ½Ñ ÑпоÑобов Ð¸Ñ ÑеализаÑии пÑименимо к ÑвеÑÑоÑнÑм нейÑоннÑм ÑеÑÑм клаÑÑиÑикаÑии изобÑажений. Ð ÑезÑлÑÑаÑе пÑоведеннÑÑ ÑеÑÑов и ÑкÑпеÑименÑов Ñделан вÑвод о наиболее Ð¿Ð¾Ð´Ñ Ð¾Ð´ÑÑÐ¸Ñ Ð½Ð°Ð¿ÑавлениÑÑ Ð¿ÑÐ¸Ð¼ÐµÐ½ÐµÐ½Ð¸Ñ Ð¸ÑÑледÑемого алгоÑиÑма в ÑвеÑÑоÑнÑÑ Ð½ÐµÐ¹ÑоÑеÑÑÑ ÐºÐ»Ð°ÑÑиÑикаÑии изобÑÐ°Ð¶ÐµÐ½Ð¸Ñ Ð¸ далÑнейÑÐ¸Ñ Ð¿ÐµÑÑпекÑивнÑÑ Ð½Ð°Ð¿ÑавлениÑÑ Ð¸Ñ Ð¸ÑполÑÐ·Ð¾Ð²Ð°Ð½Ð¸Ñ Ð² даннÑÑ ÑеÑÑÑ . The subject of the graduate qualification work is "Tuning small-sized convolutional neural networks for image classification using a genetic algorithm". This work is carried out in order to study the application of genetic algorithms in convolutional neural networks of image classification, in particular, the application in the following areas is considered: - tuning neural network hyperparameters, - training a completely untrained network, - fine tuning neural network, - selection of model architectures.In addition, a variant of the combination of the most well-proven variants of the application of the genetic algorithm is proposed. During the implementation, the basic principles of the construction and operation of convolutional neural networks were considered, an overview of existing architectures of convolutional neural networks designed for image classification was made. The review of genetic algorithms and currently existing methods of their implementation is also carried out applicable to convolutional neural networks of image classification. As a result of the conducted tests and experiments, a conclusion was made about the most suitable directions of application of the algorithm under study in convolutional neural networks of image classification and further promising directions of their use in these networks. |
Databáze: | OpenAIRE |
Externí odkaz: |