Machine Learning for Physical Scientists

Special Topic in Solid State Physics: Machine Learning for Physical Scientists

รหัสวิชา 2304641 ปีการศึกษา 2563

Announcement

• [01-17-2021] Readings for week 1-2: Mehta's section 1-8. Preliminary homework 0 here (due Friday January 22) requires you to install Jupyter notebook and relevant packages such as sklearn on your local computer. You'll get your hands dirty and get intuition why Machine Learning can be difficult.

• [01-14-2021] Unless announced here or on the facebook group, we will virtually meet every Monday from 11.00am-noon on Zoom ( Password: MLPhys ). In addition to the weekly virtual meeting, attendances are expected to learn privately at their convenience from the videos to be posted in the announcement ( 2-hour weekly lectures ). The purpose of the virtual meeting is to stimulate the discussion among attendances, as well as to have an interactive session in which questions or comments about lectures and problem sets shall be discussed.

Course Description:

In the age of Big Data, “Artificial Intelligence (AI) is the new electricity” is perhaps not an overstatement. Well-developed AI can turn seemingly useless pile of information into useful knowledge, which in turn can influence important decision making or can even fuel scientific discoveries. In the style that is suitable for physicists, this course provides fundamental concepts and techniques of Machine Learning (ML), a subfield of AI that harnesses the power of computation, whether classical or quantum, to turn data into useful computational models. We will first cover the core principles of statistical learning theory, which is a backbone of ML, including overfitting, regularization, bias-variance tradeoff, generalization, and model complexity. We will then cover important classical models of supervised and unsupervised learning such as ensemble models, Deep Learning, clustering and manifold learning, energy-based models such as Restricted Boltzmann Machines, and variational inference. Throughout the course, we will make an emphasis on natural connections between ML and statistical physics. Also, some quantum-inspired algorithms for ML will be presented towards the end of the course. In addition to pencil and paper style homework, we will also have a python-based or Tensorflow-based programming component of homework. We will use Jupyter notebooks for the programming components, where you will learn how to deploy ML algorithms in practice from physics-inspired datasets such as the Ising Model, the XY Model and topological phase transitions, Monte Carlo simulations for some scattering experiments in LHC, and etc. We will end the course with project-based presentations, where students will tackle some of open problems in ML that physicists might be able to contribute, or applying ML to solve complex physics problems of interests.

Grading Scheme:

• 50% Homework
• 20% Exam
• 30% Final Project and Presentation

References:

1. David Mackay's Information Theory, Inference, and Learning Algorithms.
2. Pankaj Mehta, et. al. A high-bias, low-variance introduction to Machine Learning for physicists
3. Caltech’s Learning from Data
4. Peter Wittek’s Quantum Machine Learning: What Quantum Computing Means to Data Mining
5. Roman Orus’ A Practical Introduction to Tensor Networks: Matrix Product States and Projected Entangled Pair States

Prerequisites:

1. Quantum Mechanics 2 (Chula Playlist)
2. Statistical Physics (Chula Playlist)
3. Python programming (Recommended Course)