Kuşaklararası etkileşimli yapay sinir ağları

Deep learning has become increasingly prevalent across diverse fields, driven by advanced learning techniques and complex network architectures such as transfer learning and teacher-student models. Transfer learning aims to achieve high performance in a target domain by leveraging knowledge from a related source domain. In contrast, teacher-student models distill knowledge from a large, complex teacher model to a smaller student model, allowing for reduced complexity without significantly compromising accuracy. Convolutional Neural Networks (CNNs), widely used in image recognition, play a key role in such models due to their efficiency and performance. However, due to their complexity, CNNs often demand substantial computational resources and long training times. This study introduces a novel hybrid neural network topology: Intergenerational Interaction Neural Networks (IINNs). The proposed hybridization topology is called Intergenerational Interaction Neural Networks. The hypothesis behind this method is that "the presence of a guiding father model enables the son model to succeed quicker and better than the others". This philosophy can be extended by incorporating additional ancestors, such as grandfathers and great-grandfathers. Unlike traditional teacher-student models, IINNs employ a pre-trained ancestor model (father) that remains static during training but actively guides the learning of the Son model. Specifically, a Self-Organizing Map (SOM) acts as the pre-trained father, and a Differential Convolutional Neural Network (DiffCNN) functions as the son. The SOM's outputs are integrated into the DiffCNN's training, enhancing convergence speed and accuracy while reducing convolutional complexity. The proposed model was evaluated on six datasets: MNIST, FashionMNIST, Birds, STL10, CIFAR-100, and CIFAR-10. It achieved superior accuracy scores of 98.58%, 96.53%, 87.49%, 86.99%, 86.78%, and 81.65%, respectively, outperforming state-of-the-art CNN, DiffCNN, and Deep Convolutional SOM models. Moreover, the model demonstrated faster convergence up to 84%, reaching 85% accuracy on more complex datasets, such as CIFAR-10, Birds, and CIFAR-100, within the first 7 to 10 epochs. At the same time, it maintained strong performance across simpler datasets like FashionMNIST, where it reached 90% accuracy by the 7th epoch, resulting in a 74% faster convergence. These results underscore the effectiveness and versatility of IINNs in accelerating training, faster convergence, and improving performance across simple and complex datasets, making them suitable for applications in medical imaging, automotive systems, and real-time scenarios like autonomous driving. Keywords: Deep learning; Convolutional neural networks; Self-organizing maps; Differential Convolutional neural networks and Image datasets.