Content
Some of the technical/mathematical concepts covered during the course are:
- Deep networks, linear and logistic regression, deep network activation functions, loss functions, optimisation by stochastic gradient descent and other methods
- Convolutional networks for image pattern recognition, using transfer learning and data augmentation to improve efficiency
- Text processing and analysis with word embeddings
- Recurrent neural networks (LSTM, GRU) to capture structures in sequential data
- Autoencoders and Generative Adversarial Networks (GANs) for mapping input data to latent space
- Object recognition and YOLO algorithm, facial recognition
- Markov decision processes, value and Q functions and optimisation
- Using deep networks as function approximators in reinforcement learning
Practical work will be based on frameworks available in Python environment, such as Numpy and Keras.
Learning objectives
The course will provide students with an overview of recent developments in artificial intelligence with emphasis on machine learning and reinforcement learning. We will study especially deep learning, and its applications in areas such as predicting from multidimensional data, performing image recognition, text analysis, and image and text generation. Deep learning will also be applied in reinforcement learning to understand how we can construct agents that interact with their environment in an intelligent manner by, for example, learning to play Atari games at a level that surpasses human performance.
Prerequisites
No prerequisites
Teaching methods
Lectures, programming exercises, essays
Learning material and recommended literature
Lecture slides, links to more detailed discussion/books/other materials to be found in slides
Evaluation criteria
Two exams: 80% of total points, assignments: 20%. Tentatively 40 points out of 100 needed for grade 1, grade increases linearly as points increase.