Institute of Data Science PhD Program
PhD-Teach-PhD Workshops - 26-31 July 2021
As part of IDS’ efforts to encourage the use of data science in research to the broader university community, we are organizing a series of PhD-Teach-PhD workshops on selected data science topics of interest. The workshops will be led by IDS’ PhD students to acquaint PhD students from other disciplines on the current topics of interests in data science.
There were 5 workshops in total, 1 2-days' workshop, and 4 1-day workshops. These workshops were open to current NUS graduate students only and were conducted in Zoom meetings. Following are the workshops details and also recordings of the workshops for those who are interested in the topics.
Workshop 1: Machine Learning for Healthcare
Presenters: Ronald Wihal Oei & Hanae Camille Carrie
Date & Time: 26-27 July 2021, 9am-5pm
Abstract: Machine learning is playing a rapidly increasing role in healthcare applications. By analysing data regarding patient hospital stays, patient visits to clinics, etc., medical practitioners can diagnose patients more accurately, can predict their future health, and come up with better treatment options. In this workshop, we will discuss the current healthcare systems, the fundamentals of machine learning for healthcare, and the challenges in applying machine learning for healthcare. This workshop will also give practical experience in applying machine learning to real-world problems in healthcare.
Recordings:
Day 1 - Part 1 (Morning Session)
Day 1 - Part 2 (Afternoon Session)
Workshop 2: Multiple Instance Learning
Presenters: Christopher Hendra
Date & Time: 28 July 2021, 9am-5pm
Abstract: In recent years, there has been a surge in the application of machine learning techniques to solve real world problems driven mainly by the exponential increase in computing power and the abundance of data labels. Nevertheless, there are many situations whereby labels are not readily available at the necessary resolutions to apply off-the-shelf machine learning algorithms, and this is especially prominent in the sciences, where labels are often obtained through imprecise experiments. To illustrate this, consider an image of a cat - it contains both groups of pixels that are relevant and irrelevant to our idea of a cat, and annotations on pixel levels are often not available. In this workshop, we will study the concept of Multiple Instance Learning (MIL), a learning paradigm that deals with tasks involving a bag of instances (partial cat images in the analogy) with labels that are available only on the bag level and not on the instance level (pixels of an image in the analogy). We will go through the various assumptions that can be made in the MIL domain and explore the various methods that years of research have produced in several fields of applications.
Recordings:
Workshop 3: Approaching Science with AI: An Introduction to Scientific Machine Learning
Presenters: Guilherme Zagatti
Date & Time: 29 July 2021, 9am-5pm
Abstract: Dynamical systems are a common way of describing phenomena in science. From harmonic oscillators, infectious diseases, and ecology to market dynamics, all of these phenomena can be described with sets of ordinary equations that detail how the variables in the system change with time and affect one another. To develop good models that explain and extrapolate empirical data, scientists may require many observations and good intuition, as well as a deep understanding of their field of expertise. Conversely, machine learning poses an alternative approach to research: data in, predictions out. This black-box approach often eschews generalization and leaves little room for explanation and understanding. Recently, a new paradigm that blends scientific methods with machine learning called scientific machine learning has emerged. Using the fact that neural networks are universal function approximators, we can blend neural networks with ordinary differential equations to facilitate explainable model discovery from data. In particular, a new breed of neural network called NeuralODE can help us make sense of non-linear dynamics. This course will introduce scientific machine learning using the Julia programming language and its SciML environment.
Recordings:
Workshop 4: Speech Processing with Deep Learning
Presenters: Pan Zexu
Date & Time: 30 July 2021, 9am-5pm
Abstract: One of our most important faculties is our ability to listen to, and follow, one speaker in the presence of others. This is such a common experience that we may take it for granted; we may call it ‘the cocktail party problem.’ Since this problem was formulated, the quest for equipping machines with such an ability has never stopped. With the advent of deep learning, speech separation and speaker extraction represent methods for solving the cocktail party problems through engineering approaches. These techniques have facilitated many speech processing tasks in multi-talker acoustic environments including automatic speech recognition, speaker verification and hearing aid development. In this workshop, we will provide a brief introduction to the cocktail party problem and recent deep learning based techniques for solving it.
Recordings:
Workshop 5: Deep Generative Modelling from the Perspective of Divergence Measures
Presenters: Li Shen
Date & Time: 31 July 2021, 9am-5pm
Abstract: Generative modelling is a highly active research topic based on explicitly or implicitly modelling input data with high accuracy (e.g., images or audio signals). This topic promises a wide range of applications including data generation, variational inference, and density estimation. Recent advances have included the highly successful Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), and more. In this workshop, we will introduce these advances in deep generative models through the lens of their objective functionals—divergence between probability distributions—and then present their connections, strengths and weaknesses. We further illustrate several unsolved issues and possible research directions.
Recordings: