In machine and deep learning, a convolutional neural network (CNN) is a powerful tool that can take in an input image, assign importance (learnable parameters) to various aspects/objects in the image and be able to discriminate one from the other. The architecture of a CNN is similar to that of the connectivity schema of neurons in the brain of human and was inspired by the structure of the visual cortex. The pre-processing step in a CNN is much lower as compared to other machine learning algorithms. Moreover, CNN has the ability to automatically extract the features and representations, and learn the intermediate filters, while in most of the machine learning methods the filters and features are generated by hand with variety of challenges.
The Fashion-MNIST clothing classification problem is a new standard dataset used in computer vision and deep learning. The dataset contains small square gray-scale images of 10 types of clothing, such as sneaker, pullover, coat, and more. Since the dataset has good match to MNIST, it can be used as the benchmark for evaluating the deep convolutional neural networks for image classification. Chest x-ray dataset consists of synthetic chest X-ray images. Images are generated using an unsupervised domain adaptation approach by leveraging class conditioning and adversarial training from source datasets.
In this assignment, we are to analyze the performance of deep models in facing with various problems. We design the deep model with dense and convolutional layers to compare their performance in various conditions. Also, we evaluate the effect of pretrained models on chest x-ray dataset. Our experiments demonstrate that the convolutional models show better performance in learning the general and specific patterns of problem. Moreover, the pretrained models outperform the classic models with transferring knowledge across domains.