Résultats de recherche pour « CNNs »

We extend Memory-modulated Convolutional Neural Networks by introducing stochastic memory decay through multiple functions-linear, exponential, and logarithmic-replacing the earlier linear-only degradation. Each memory tensor undergoes random decay per inference, and its temporal distance is inferred via the relative change in mean activation. This approach enables the modeling of subjective time, where memories created simultaneously diverge in their perceived age due to differing decay paths. These results demonstrate that memory degradation can simulate temporal perception in CNNs, offering a biologically inspired foundation for time-aware neural computation.

Convolutional neural networks (CNNs) achieve strong performance in visual recognition tasks but often lack interpretability, while existing post-hoc explainable artificial intelligence (XAI) methods introduce additional computational overhead and separate explanation pipelines, limiting their practicality in resource-constrained environments. To address this limitation, this work proposes a lightweight self-explainable framework, termed Feature-Activated Convolutional Neural Network (FA-CNN), which enables simultaneous prediction and explanation through minimal architectural modification of conventional CNNs. FA-CNN generates class-agnostic Feature Activation Maps (FAMs) as intermediate representations for classification and subsequently transforms them into class-specific saliency maps using the learned classifier parameters, providing intrinsic explanations without requiring external XAI methods. An interpretation-aware training strategy is further introduced, where the generated saliency maps are guided by an auxiliary supervision loss derived from automatically generated proxy spatial guidance within a weakly supervised learning framework. Experiments on four benchmark datasets demonstrate that the proposed models achieve classification performance comparable to their standard CNN counterparts while producing faithful and visually meaningful explanations. Quantitative evaluation using multiple saliency assessment metrics shows competitive explanation quality compared with representative post-hoc methods, including Grad-CAM, Grad-CAM++, and LayerCAM. Although FA-CNN introduces a modest computational overhead during training, it generates class predictions and saliency maps simultaneously during inference, enabling substantially faster explanation generation. The source code is made publicly available at https://github.com/Sayankumar007/LINK-LINK-LINK.

Our paper presents some creative advancements in the image in-painting techniques for small, simple images for example from the CIFAR10 dataset. This study primarily targeted on improving the performance of the context encoders through the utilization of several major training methods on Generative Adversarial Networks (GANs). To achieve this, we upscaled the network Wasserstein GAN (WGAN) and compared the discriminators and encoders with the current state-of-the-art models, alongside standard Convolutional Neural Network (CNN) architectures. Side by side to this, we also explored methods of Latent Variable Models and developed several different models, namely Pixel CNN, Row Long Short Term Memory (LSTM), and Diagonal Bidirectional Long Short-Term Memory (BiLSTM). Moreover, we proposed a model based on the Pixel CNN architectures and developed a faster yet easy approach called Row-wise Flat Pixel LSTM. Our experiments demonstrate that the proposed models generate high-quality images on CIFAR10 while conforming the L2 loss and visual quality measurement.

Brain tumors are a serious illness that needs to be diagnosed as soon as possible. Convolutional neural networks (CNNs) can automate parts of medical image analysis, as shown by recent developments in deep learning. This could help clinicians and speed up diagnosis. In this work, we describe a fully reproducible pipeline that uses transfer learning to classify brain tumors into four categories: glioma, meningioma, pituitary, and healthy. We make use of an MRI dataset with over 7,000 labeled images that is made publicly available by Mendeley Data. To produce four classes, our method uses a ResNet-18 model that was pretrained on ImageNet and has had its last layer swapped out. After normalizing and resizing each MRI scan to 224 x 224 pixels, standard training methods (such as cross-entropy loss and Adam optimizer) are used. Instead of emphasizing particular outcomes, we emphasize reproducibility and a clear methodology, providing examples of data preprocessing, model setup, and sample visualizations. This framework shows how pre-trained CNNs can be effectively modified to classify brain tumors, cutting down on training time without sacrificing performance.