Applied Machine Learning

About this course

Machine learning has become a fundamental part of many research and commercial applications. Modern frameworks like Keras/Tensorflow and Scikit-learn have made it easy to write programs that implement sophisticated machine learning algorithms that can be readily applied to text, image and audio data. This course offers an introduction to the fundamentals of machine learning and quickly scales to applications that include data from image and text to audio and music. The student is expected to obtain a clear understanding of the basic mathematical principles and working knowledge on developing end-to-end research implementations, i.e., from data curation to proper presentation of results. All programming is in Python using the Keras framework, while the course offers material from books, slides and GitHub code. Aim of this course is to provide a clear understanding of the basic principles that underly modern machine learning algorithms, while enabling end-to-end implementation of basic versions of modern machine learning systems.

Expected learning outcomes

On completion of this course, students will be able to: – Identify the correct class of algorithms for solving specific problems. – Prepare data in optimal representations for specific tasks. – Follow steps to mitigate common obstacles concerning overfit and error stagnation. – Present results in a scientifically rigorous manner.

Indicative Syllabus

    The course starts with basic ideas around data dimensionality, dimensionality reduction and linear regression. This study leads to a straightforward understanding of artificial neural networks (ANNs). The MNIST digits dataset is employed for introducing basic supervised learning with feedforward ANNs and towards the introduction of convolutional layers. 1D convolutional layers are introduced for waveforms and 2D for spectrograms. Autoencoders are then implemented which lead to Variational Autoencoders and Generative Adversarial Networks, initially with the MNIST dataset and then for audio. Recurrent Neural Networks (LSTM and GRU) are also examined for sequence learning and sequence-to-sequence translation, along with Temporal Convolutional Networks (TCNs). This study leads to Transformer encoder-decoder architectures (under the hood of ChatGPT), which are understood deeply and developed with ready-made implementations of KerasNLP.

    Teaching / Learning Methodology

    ΤΒΑ

    Recommended Reading

    [1] Géron, A. (2022). Hands-on machine learning with Scikit-Learn, Keras, and TensorFlow. ” O’Reilly Media, Inc.”. [2] Foster, D. (2022). Generative deep learning. ” O’Reilly Media, Inc.”.

    Prerequisites

    Start Date

    TBA

    End Date

    TBA

    Apply

    TBA

    ENROLL NOW

    Local Course Code

    0807.8.004.1

    Cycle

    TBA

    Year of study

    TBA

    Language

    English

    Study Load

    Lectures 2 hours per week, Practiacal 2 hours per week,  In total 4 hours per week  6 ECTS

    Mode of delivery

    TBA

    Instructors

    Dr. Maximos Kaliakatsos

    Course coordinator

    Dr. Maximos Kaliakatsos

    E-mail

    maximoskp@hmu.gr