Abstract Nowadays, transfer learning has shown promising results in many applications. However, most deep transfer learning methods such as parameter sharing and fine‐tuning are still suffering from the lack of parameters transmission strategy. In this paper, we propose a new optimization model for parameter‐based transfer learning in convolutional neural networks named STP‐CNN. Indeed, we propose a Lasso transfer model supported by a regularization term that controls transferability. Moreover, we opt for the proximal gradient descent method to solve the proposed model. The suggested technique allows, under certain conditions, to control exactly which parameters, in each convolutional layer of the source network, which will be used directly or adjusted in the target network. Several experiments prove the performance of our model in locating the transferable parameters as well as improving the data classification.

Abstract This paper introduces an efficient real-coded genetic algorithm (RCGA) evolved for constrained real-parameter optimization. This novel RCGA incorporates three specially crafted evolutionary operators: Tournament Selection (RS) with elitism, Simulated Binary Crossover (SBX), and Polynomial Mutation (PM). The application of this RCGA is directed toward optimizing the MLPRGA+5 model. This model is designed to configure Multilayer Perceptron neural networks by optimizing both their architecture and associated hyperparameters, including learning rates, activation functions, and regularization hyperparameters. The objective function employed is the widely recognized learning loss function, commonly used for training neural networks. The integration of this objective function is supported by the introduction of new variables representing MLP hyperparameter values. Additionally, a set of constraints is thoughtfully designed to align with the structure of the Multilayer Perceptron (MLP) and its corresponding hyperparameters. The practicality and effectiveness of the MLPRGA+5 approach are demonstrated through extensive experimentation applied to four datasets from the UCI machine learning repository. The results highlight the remarkable performance of MLPRGA+5, characterized by both complexity reduction and accuracy improvement.

Abstract Time Series Classification (TSC) started getting a lot of attention recently, mostly due to the important real‐world applications of time series such as in the financial industry. In this context, various approaches have been proposed to treat this important and challenging problem in data mining. These approaches include deep learning models that are outperformed the traditional classification techniques. However, the choice of an optimal and efficient deep neural network for TSC is still a major problem. Motivated by this challenge, we propose in this paper, an optimized deep learning framework for TSC tasks, using normal cloud representation and convolutional neural networks optimization (NCR‐CNNO). Our approach consists of two phases: In the first one, we convert the raw time series to a matrix of characteristics using two‐dimensional normal cloud representation. In the second phase, we suggest a CNN to handle the matrix generated in the first phase, as well as, an optimization model is proposed to optimize the unnecessary and redundant parameters. Experiments conducted on extensive time series datasets demonstrate that NCR‐CNNO yields significant improvement in the performance of time series classification compared to the state of the art.