Deep Learning

The ability of computerized systems to acquire vast amounts of data and display them in informative ways raises our expectations for fast, accurate identification or recognition of events or objects and for predictions about future events. Machine learning and artificial intelligence (AI) have fulfilled those needs to some degree. Over the last 10 years, a versatile architectural style of artificial neural networks called deep learning has emerged as the most promising answer to those expectations. Today, deep learning is the primary technique for analysis and resolution of many issues in data analyses and natural sciences, linguistics, and engineering. We use deep learning for image classification, manipulation and generation, speech recognition and synthesis, natural language translation, sound and music manipulation and generation, navigation of self-driving cars, and many other activities. In this course, students master several key architectures for implementation of deep learning networks, such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), long short-term memory networks (LSTMs), autoencoders, generative adversarial networks (GANs), transformers with attention, and graph neural networks. We provide references to many practical applications where those architectures are successfully used. The course starts with a review of the theoretical foundations of the neural networks approach to machine learning including auto-differentiation and backpropagation. The emphasis of the course is on practical applications of deep learning using Keras (packages within TensorFlow 2.x framework) and PyTorch.