Mastering neural networks using TensorFlow and Keras offers a robust and user-friendly approach to designing, implementing, and optimizing deep learning models. TensorFlow, developed by Google, is a comprehensive open-source library for numerical computation and machine learning, offering extensive capabilities for building and deploying neural networks. Keras, an Application Programming Interface (API) within TensorFlow, simplifies the construction of neural networks with its user-friendly, high-level interface, making it accessible for both beginners and experienced users.

This paper presents a laboratory exercise designed for the Computer Vision course within the Multimedia Masters Program. The exercise is based on the Python programming language and the TensorFlow library, with examples utilizing the Keras API. It begins by introducing several simple neural networks with fully connected (dense) layers. Following this, a convolutional neural network (CNN) is introduced for the classification of MNIST handwritten digits. The exercise then covers a denoising autoencoder, which is used to denoise MNIST images corrupted with Gaussian noise. Finally, a generative adversarial network (GAN) is presented, trained on the MNIST dataset to generate new images of handwritten digits. Throughout the laboratory session, students may encounter various challenges, particularly due to different levels of background knowledge, especially in mathematics. Additionally, it may be important to develop lightweight models that can be trained on a standard computer without the need for an expensive GPU, or alternatively, provide pre-trained model weights.

An analysis will then be conducted, examining the total number of M.Sc. theses, both completed and in progress, under the same supervisor in the Multimedia Masters Program over five academic years: 2019/2020 to 2023/2024. The analysis will focus on M.Sc. theses from two courses: Computer Vision and Multimedia Video Technology. The findings suggest that, in recent years, all theses proposed within the Computer Vision course have incorporated neural networks.

In conclusion, TensorFlow and Keras enable students to close the knowledge gap between theory and practice in educational and professional development programs. It gives them a deeper and more useful understanding of deep learning in fields like computer vision and natural language processing.

Keywords: Neural networks, Fully connected layer, CNN, Autoencoder, GAN, TensorFlow, Keras.